Pith. sign in

REVIEW 5 major objections 4 minor 288 references

This thesis claims that injecting geometric priors—skyline and ground-plane cues, SE(3)-equivariant surfels, focal-stack attention, and wavelet depth features—lets deep models consistently outperform prior state-of-the-art on camera pose es

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A PhD thesis showing that adding geometric priors (skyline, normals, focus cues, wavelet depth) to deep networks improves pose estimation, registration, depth prediction, and reconstruction.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Four methods, three already published, wrapped in a thesis whose abstract overpromises and whose integration claim is never tested. the 5 major comments →

arxiv 2509.01873 v1 pith:QBPOGJN3 submitted 2025-09-02 cs.CV cs.AI

Doctoral Thesis: Geometric Deep Learning For Camera Pose Prediction, Registration, Depth Estimation, and 3D Reconstruction

Xueyang Kang This is my paper
classification cs.CV cs.AI
keywords geometric deep learningcamera pose estimationpoint cloud registrationfocal stack depth estimationimplicit signed distance fieldswavelet transformSE(3) equivariance3D reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The thesis tries to show that deep learning models for core 3D vision tasks become more accurate and robust when traditional geometric constraints are built into the architecture or the training signal. It proposes four hybrid methods: drone camera orientation tracking anchored by skyline and ground-plane cues fused with IMU data by an adaptive particle filter; point cloud registration via SE(3)-equivariant features on 2D surfels; depth estimation from focal stacks of arbitrary length using a Transformer with a latent LSTM; and implicit SDF reconstruction conditioned on wavelet-transformed depth features. The author claims each method outperforms state-of-the-art baselines on public benchmarks in its task, with particular gains in low-overlap registration, arbitrary-length focal stacks, and fine geometric detail. A sympathetic reader would care because the thesis positions the four methods as interchangeable modules of a larger reconstruction pipeline rather than isolated results.

Core claim

On the paper's own terms, geometric deep learning bridges the gap between traditional geometry-based methods and data-driven approaches: each of the four proposed methods consistently outperforms state-of-the-art techniques on its individual 3D vision task. The camera pose system uses natural skyline and ground-plane primitives, fused with IMU data through an adaptive particle filter on a spherical manifold, to achieve real-time orientation tracking that is more stable than IMU-only solutions. The registration method introduces a surfel-based SE(3)-equivariant framework that learns position and orientation representations from 2D Gaussian surfels, achieving superior accuracy in indoor and ou

What carries the argument

Four mechanisms carry the argument: (1) natural geometric primitives—the skyline and ground plane—used as reference cues for roll and pitch tracking, fused by an adaptive particle filter sampling on a multi-resolution spherical manifold; (2) SE(3)-equivariant 2D Gaussian surfel features, encoded by an E2PN-style symmetric convolutional kernel, to make registration robust to rigid transformations and input uncertainty; (3) a Transformer self-attention encoder coupled with an LSTM that sequentially fuses tokens along the focal stack dimension, enabling arbitrary stack lengths and monocular pre-training; and (4) wavelet-transformed depth features, projected onto triplanes and fused via a 2D U-N

Load-bearing premise

The pose-tracking system assumes the skyline can be approximated by a straight line and that, above 300 meters, the ground is flat and perpendicular to gravity; the thesis itself notes that curved mountain skylines challenge this assumption.

What would settle it

Collect a flight sequence in terrain with a curved skyline and uneven ground below 300 meters, using servo-motor ground truth for roll and pitch; if the fusion's RMSE stops beating the IMU-only Madgwick filter, the central drift-resistance claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Camera orientation tracking runs in real time on an embedded Jetson Nano and beats IMU-only filtering against drift in natural scenes, making it deployable for drone-based imaging.
  • Surfel-based equivariant registration retains high recall on low-inlier point clouds and degrades gracefully even with only 256 sampled points, supporting robotics and mobile scanning use cases.
  • FocDepthFormer accepts any number of focal-stack frames, so one trained model can serve different stack sizes and can be pre-trained on monocular RGB-D data to reduce reliance on scarce focal-stack datasets.
  • Wavelet-conditioned implicit SDF reconstruction yields more complete and detailed meshes across object-scale and architectural-scale scenes, and is described as a plug-and-play module for existing implicit SDF frameworks.
  • The modular design suggests that improvements in pose, registration, depth, and reconstruction each strengthen the corresponding stage of a full 3D reconstruction pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the pose estimator's reliance on a single straight skyline and a flat ground plane means the claimed robustness likely holds only in open terrain; the thesis itself acknowledges curved mountain skylines as a failure mode.
  • Editorial inference: because the four modules are evaluated separately, the thesis does not prove that chaining them into one end-to-end reconstruction pipeline preserves the individual gains; a system-level integration test would be a natural next step.
  • Editorial inference: the wavelet-conditioning idea could plausibly transfer to other implicit representations, such as radiance-field or Gaussian-splatting reconstructions, where high-frequency detail is also a bottleneck.
  • Editorial inference: the latent LSTM fusion for focal stacks suggests a general recipe for any ordered image set with per-frame focus cues, including smartphone refocus stacks or light-field captures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The thesis combines geometric priors with deep learning across four 3D vision tasks. Chapter 3 proposes a camera-orientation tracker that fuses IMU data with skyline and ground-plane cues through an adaptive particle filter on a spherical manifold, targeting UAV/natural-environment operation. Chapter 4 introduces a surfel-based registration network with an SE(3)-equivariant encoder, cross-attention, and Huber loss, evaluated on 3DMatch and KITTI. Chapter 5 presents FocDepthFormer, a Transformer-LSTM architecture for depth-from-focal-stack with arbitrary stack length, evaluated on DDFF, FOD500, Mobile Depth, and LightField4D. Chapter 6 proposes a wavelet-conditioned implicit SDF reconstruction method that fuses multi-scale wavelet-transformed depth features with triplane features, evaluated on DTU, Tanks and Temples, and a cultural heritage dataset. The abstract claims that the four methods 'consistently outperform state-of-the-art techniques across individual 3D vision tasks.'

Significance. If fully supported, the thesis would contribute practical geometry-aware modules for pose tracking, registration, depth estimation, and reconstruction. Notable strengths are the real-time embedded deployment of the pose tracker, the use of explicit surface orientation and uncertainty in registration, the LSTM-based variable-length focal-stack processing, and the wavelet-conditioned implicit surface prior. However, the broad SOTA claim is not consistently supported by the reported tables, and one chapter's assumptions are untested in the environments the chapter targets. The correctable nature of these issues makes the manuscript worth revising.

major comments (5)
  1. [Abstract and §4, Table 4.1] The abstract states that the four methods 'consistently outperform state-of-the-art techniques across individual 3D vision tasks.' Table 4.1 directly contradicts this for Chapter 4 on KITTI: Ours has RE=1.57°, while SpinNet achieves RE=1.08°. The text in §4.4.1 claims a 'smaller rotation error (around 8-11% reduction)' compared to the second-best model, but on KITTI the rotation error is about 45% larger than SpinNet's. This is load-bearing for the central claim and must be corrected, either by reporting the comparable/competitive status on that metric or by removing the overclaim.
  2. [§3.4, Tables 3.1 and 3.2] The mixed-test duration is reported as 960 s in Table 3.1 and 9200 s in Table 3.2. Several entries in Table 3.2 are missing (notably R-VIO and DM-VIO in several rows), yet the prose states that the proposed method 'performs consistently across different sequences' and compares favorably. In addition, the text first says ORB-SLAM3, R-VIO, and DM-VIO were selected for comparison, then says these algorithms were found unsuitable and the comparison is against IMU-only Madgwick filtering. The experimental protocol and table completeness must be clarified before the robustness and superiority claims of Chapter 3 can be accepted.
  3. [§3.3.2, §3.3.3, §3.5] The skyline tracking assumes a straight-line skyline (Eqs. 3.1–3.4) and the ground-plane cue assumes a flat ground orthogonal to gravity above 300 m (Eqs. 3.8–3.12). Section 3.5 explicitly acknowledges that 'complex terrains with curved mountains challenge the straight-line assumption.' Yet all experiments in §3.4 are rooftop or drone sequences with largely linear horizons and planar ground. There is no evaluation on curved mountain skylines, partially occluded horizons, or non-planar terrain, which are precisely the 'natural environments' claimed in the abstract. If these geometric cues fail, the particle filter is left with only IMU, and the claimed advantage over IMU-only solutions vanishes. This is a correctness-risk concern: the central pose-tracking claim for unstructured natural scenes is not yet demonstrated.
  4. [§5.3.1, Table 5.3] In the FOD500 results, AiFNet is reported with absRel=0.400, sqrRel=4.32e-1, yet δ1=85.12. These values are mutually implausible: an absRel of 0.400 typically corresponds to much lower δ1 accuracy. This suggests a data entry or metric-calculation error. Because the table is used to show that the proposed method outperforms baselines, the correctness of the baseline numbers is essential. Please verify all entries in Table 5.3 and report corrected values.
  5. [§4.4.2, Table 4.4] The robustness analysis states that 'the registration error remains within 4-5% of the input translation, and the rotation error within 3-4%.' For the largest perturbation (100°, 200 cm), Table 4.4 reports TE=12.59 cm, which is 6.3% of the input translation, not within 5%. The textual summary is therefore inconsistent with the table. Either the claim or the table should be corrected.
minor comments (4)
  1. [§3.3] The equation numbering is inconsistent: the text refers to 'Equation 1' while equations are numbered (3.1)–(3.20). Algorithm 1 has garbled line references and undefined symbols (ε, δt, b, Ω1, Ω2, Ω3). Please polish the pseudocode and define all variables.
  2. [§5.2.1 and §5.2.4] The Circle-of-Confusion definition appears both as Eq. (5.1) and Eq. (5.9), with slightly different notation. The symbol δ is used both for the accuracy threshold in Eq. (5.15) and for the variance in Eq. (5.8), causing ambiguity.
  3. [§4, Tables] In Table 4.1, the RoReg row contains a stray arrow '↓' and the column alignment is inconsistent. In Table 4.2, the 'Average' column is formatted differently from the other columns. Please unify table formatting.
  4. [Overall] There are several typos and formatting artifacts, e.g., 'AutoV AE' in §6.10, '3D Unet' without hyphenation, and inconsistent use of 'SOTA' vs. 'state-of-the-art.' A careful proofreading pass is needed.

Circularity Check

0 steps flagged

No significant circularity; the thesis validates four independent methods against external benchmarks without reducing any derivation to its own inputs.

full rationale

The thesis is an empirical compilation. Each content chapter proposes a method and evaluates it on public benchmarks (3DMatch/KITTI, DDFF/FOD500, DTU/Tanks-and-Temples). The geometric relationships used in Chapter 3 (skyline slope/intercept to roll/pitch, ground-plane normal via Rodrigues' formula) are analytical cues, not fitted quantities renamed as predictions. The particle filter variances are tuned constants, not the claimed output. Chapter 4's equivariant encoder is an architectural choice adapted from prior work and tested by ablations; no prediction is defined in terms of the fitted parameters. Chapter 5 uses the Circle-of-Confusion formula only as motivation, while depth is learned from data with external supervision. Chapter 6 conditions an implicit SDF on wavelet-transformed depth features from a monocular depth prior; although this raises a possible domain-overlap concern, the text does not assert that the prior was trained on the evaluation targets, and conditioning a reconstruction on an independent depth cue is not definitionally circular. The self-citations in Chapters 3–5 refer to the author's own peer-reviewed papers, which are independent prior publications and are not invoked as the sole justification for the present claims. The acknowledged limitations (curved skylines, view-dependent appearance, textureless scenes, memory footprint) are honest caveats that affect generalization but do not make any derivation equivalent to its inputs. No load-bearing step reduces by construction to a fitted value or to a self-citation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 0 invented entities

The thesis introduces no new physical entities. Its central claims rest on standard computer vision and geometry assumptions (perspective projection, Rodrigues formula, CoC model, SDF representation), plus several domain-specific assumptions about natural scenes and depth priors, which are acknowledged in the text. The main fitted components are network weights and hand-set hyperparameters, several of which are not fully specified.

free parameters (6)
  • Particle filter variances and thresholds (Chapter 3) = Not fully specified; 'set as a constant according to practical tests'
    The fusion quality in Algorithm 1 depends on variances delta_C1, delta_C2, distance threshold epsilon, and sampling resolutions Omega_1,2,3, which are chosen by hand and by practical tuning.
  • Surfel uncertainty normalization C and inverse-depth range (Chapter 4, Eq. 4.5-4.7) = Not specified numerically
    The surfel radius formula uses a normalization factor C and truncated inverse-depth range (rho_min, rho_max) that are set by the sensor model and affect registration robustness.
  • Huber loss threshold delta (Chapter 4, Eq. 4.10) = 0.6 m
    The threshold balances L1 and L2 loss behavior and is chosen for the 3DMatch/KITTI experiments.
  • LSTM token activation threshold and loss weight (Chapter 5) = activation threshold 0.4, alpha = 0.2
    The token L2-norm threshold controls how many tokens enter the LSTM, and alpha weights the sharpness regularizer in Eq. 5.8; both are tuned on validation data.
  • Wavelet decomposition levels and pretrained wavelet autoencoder weights (Chapter 6) = Three resolutions; autoencoder pretrained on sharp depth maps
    The model conditions on wavelet features from a pretrained autoencoder, whose weights are fitted on depth map data and directly influence reconstruction fidelity.
  • Pretrained monocular depth model used in Chapter 6 = Not named in the available text
    The reconstruction pipeline processes images through a monocular depth model before wavelet transform; the quality of that external prior affects the final SDF reconstruction.
axioms (7)
  • domain assumption The skyline in natural scenes can be approximated as a straight line for roll and pitch estimation (Chapter 3, Eq. 3.1-3.4)
    Used to fit slope and intercept via least squares; the thesis acknowledges in Section 3.5 that curved mountains challenge this assumption.
  • domain assumption The ground plane is flat and orthogonal to the gravity vector when the camera is above 300 m (Chapter 3, Section 3.3.3)
    The trigonometic back-projection in Eq. 3.8-3.11 requires a flat ground plane at known height to estimate the plane normal; this is an environmental assumption.
  • domain assumption The camera is never upside down during operation (Chapter 3, Section 3.3.2)
    Used to search for the skyline along a predicted vertical direction; inverted camera states would break the search.
  • domain assumption Surface normals derived from depth gradients are accurate enough to serve as equivariant input signals (Chapter 4, Section 4.3.1)
    The surfel representation uses normal vectors computed via a Sobel operator on depth; noisy depth maps lead to noisy normals, which the model must tolerate.
  • standard math The Circle-of-Confusion model correctly describes defocus blur as a function of depth and focus distance (Chapter 5, Eq. 5.1)
    The CoC equation is a standard optics model; the thesis uses it to motivate depth-from-defocus and to design the sharpness regularizer.
  • standard math The zero-crossing of an implicit SDF corresponds to the reconstructed surface and can be optimized via ray sampling (Chapter 6, Section 6.1)
    This is the standard formulation of implicit SDF reconstruction; the thesis relies on it for training and marching-cubes extraction.
  • domain assumption A pretrained monocular depth model provides a reliable geometric prior for wavelet feature conditioning (Chapter 6, Figure 6.7)
    The pipeline conditions on depth maps from an external monocular depth estimator; if this prior is inaccurate, the wavelet features inject wrong geometry into the SDF.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Doctoral Thesis: Geometric Deep Learning For Camera Pose Prediction, Registration, Depth Estimation, and 3D Reconstruction." pith.science (2026). https://pith.science/paper/QBPOGJN3

@misc{pith2026250901873,
  author       = {Pith},
  title        = {Pith review of: Doctoral Thesis: Geometric Deep Learning For Camera Pose Prediction, Registration, Depth Estimation, and 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBPOGJN3}},
  note         = {Machine review of arXiv:2509.01873}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Modern deep learning developments create new opportunities for 3D mapping technology, scene reconstruction pipelines, and virtual reality development. Despite advances in 3D deep learning technology, direct training of deep learning models on 3D data faces challenges due to the high dimensionality inherent in 3D data and the scarcity of labeled datasets. Structure-from-motion (SfM) and Simultaneous Localization and Mapping (SLAM) exhibit robust performance when applied to structured indoor environments but often struggle with ambiguous features in unstructured environments. These techniques often struggle to generate detailed geometric representations effective for downstream tasks such as rendering and semantic analysis. Current limitations require the development of 3D representation methods that combine traditional geometric techniques with deep learning capabilities to generate robust geometry-aware deep learning models. The dissertation provides solutions to the fundamental challenges in 3D vision by developing geometric deep learning methods tailored for essential tasks such as camera pose estimation, point cloud registration, depth prediction, and 3D reconstruction. The integration of geometric priors or constraints, such as including depth information, surface normals, and equivariance into deep learning models, enhances both the accuracy and robustness of geometric representations. This study systematically investigates key components of 3D vision, including camera pose estimation, point cloud registration, depth estimation, and high-fidelity 3D reconstruction, demonstrating their effectiveness across real-world applications such as digital cultural heritage preservation and immersive VR/AR environments.

Figures

Figures reproduced from arXiv: 2509.01873 by Xueyang Kang.

Figure 1.1
Figure 1.1. Figure 1.1: Chapter relationships and outline structures of dependencies, [PITH_FULL_IMAGE:figures/full_fig_p022_1_1.png] view at source ↗
Figure 2.1
Figure 2.1. Figure 2.1: Geometric deep learning diagram combines various deep learning [PITH_FULL_IMAGE:figures/full_fig_p027_2_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: , deep learning regression models are applied to both object pose [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗
Figure 2.2
Figure 2.2. Figure 2.2: Deep learning methods for pose regression. (a) PoseNet: a CNN [PITH_FULL_IMAGE:figures/full_fig_p028_2_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: , including the traditional ICP-based approach, multi-level feature [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 2.3
Figure 2.3. Figure 2.3: Deep learning methods for point cloud registration. (a) [240]: an [PITH_FULL_IMAGE:figures/full_fig_p029_2_3.png] view at source ↗
Figure 2.4
Figure 2.4. Figure 2.4: Deep learning architectures for depth prediction. (a) Deep Vir [PITH_FULL_IMAGE:figures/full_fig_p031_2_4.png] view at source ↗
Figure 2
Figure 2. Figure 2: depicts a fundamental Structure from the Motion approach, [PITH_FULL_IMAGE:figures/full_fig_p033_2.png] view at source ↗
Figure 2.5
Figure 2.5. Figure 2.5: Illustration of structure from motion by multi-view images, image [PITH_FULL_IMAGE:figures/full_fig_p034_2_5.png] view at source ↗
Figure 2.6
Figure 2.6. Figure 2.6: Common SLAM framework: (a) visual SLAM, (b) Lidar-based [PITH_FULL_IMAGE:figures/full_fig_p036_2_6.png] view at source ↗
Figure 2.7
Figure 2.7. Figure 2.7: 3D reconstruction by using a 3D recurrent method (a), and [PITH_FULL_IMAGE:figures/full_fig_p037_2_7.png] view at source ↗
Figure 2.8
Figure 2.8. Figure 2.8: Implicit SDF for a wide range of reconstruction problems, from [PITH_FULL_IMAGE:figures/full_fig_p039_2_8.png] view at source ↗
Figure 2.9
Figure 2.9. Figure 2.9: Neural Radiance Field (at the left) alongside its extensions: 3D [PITH_FULL_IMAGE:figures/full_fig_p040_2_9.png] view at source ↗
Figure 2.10
Figure 2.10. Figure 2.10: The NICE-SLAM framework [288] tracks ego-body pose and [PITH_FULL_IMAGE:figures/full_fig_p041_2_10.png] view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: Illustration of pose orientation estimation from image frames on [PITH_FULL_IMAGE:figures/full_fig_p047_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: Open source hardware setup. plane in the current frame with those in the reference frame. The compen￾sation angles from various pipelines are then fed into the proposed particle filter presented in the Section below to obtain fusion orientations, further as input for the controller to stabilize the camera. Video Input Binary Segmentation & Skyline Isolation Skyline Ground Plane IMU Data Geometry Tracking… view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: Block diagram of the presented orientation estimation algorithm. [PITH_FULL_IMAGE:figures/full_fig_p048_3_3.png] view at source ↗
Figure 3.4
Figure 3.4. Figure 3.4: Failure case demo using traditional OpenCV pipeline. [PITH_FULL_IMAGE:figures/full_fig_p049_3_4.png] view at source ↗
Figure 3.5
Figure 3.5. Figure 3.5: Sample images and ground truth masks for training. [PITH_FULL_IMAGE:figures/full_fig_p050_3_5.png] view at source ↗
Figure 3.6
Figure 3.6. Figure 3.6: Segmented images for skyline search. Roll and pitch have a specific tolerance range to avoid unnecessary op￾erations, so processing is only triggered when the movement is out of this range. Roll angle can be predicted from the slope m of the skyline, followed by pitch estimation, which is on top of the image result after roll compen￾sation. There is a total of three cases in our system, that is pure roll… view at source ↗
Figure 3.7
Figure 3.7. Figure 3.7: Arrows between them stand for transition conditions. In our [PITH_FULL_IMAGE:figures/full_fig_p055_3_7.png] view at source ↗
Figure 3.7
Figure 3.7. Figure 3.7: Lifetime phases of particle filter sampling on a spherical surface. [PITH_FULL_IMAGE:figures/full_fig_p056_3_7.png] view at source ↗
Figure 3.8
Figure 3.8. Figure 3.8: Simulation test setup on top of the building. [PITH_FULL_IMAGE:figures/full_fig_p057_3_8.png] view at source ↗
Figure 3
Figure 3. Figure 3: further validates the consistency of our method. The test is [PITH_FULL_IMAGE:figures/full_fig_p060_3.png] view at source ↗
Figure 3.9
Figure 3.9. Figure 3.9: The Green arrow is a mean error, the orange line is a median [PITH_FULL_IMAGE:figures/full_fig_p060_3_9.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: Different Equivariant model designs. (a) is equivariant [PITH_FULL_IMAGE:figures/full_fig_p065_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: (Left) Color image decomposed into superpixels of large size. [PITH_FULL_IMAGE:figures/full_fig_p068_4_2.png] view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: (Left) Color points cast from the depth map aligned with the [PITH_FULL_IMAGE:figures/full_fig_p069_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: The network structure features a shared encoder for surfels (6 [PITH_FULL_IMAGE:figures/full_fig_p071_4_4.png] view at source ↗
Figure 4.5
Figure 4.5. Figure 4.5: Recovering discreteized SO(3) ′ from the quotient feature S 2 ′ by permutation order. The surfel undergoes convolution with two distinct symmetric kernels, κ1 and κ2, as shown in [PITH_FULL_IMAGE:figures/full_fig_p072_4_5.png] view at source ↗
Figure 4.6
Figure 4.6. Figure 4.6: Comparison results on KITTI [66]. For each dataset, the top [PITH_FULL_IMAGE:figures/full_fig_p076_4_6.png] view at source ↗
Figure 4.7
Figure 4.7. Figure 4.7: Huber loss learning curve under different thresholds. [PITH_FULL_IMAGE:figures/full_fig_p078_4_7.png] view at source ↗
Figure 5.1
Figure 5.1. Figure 5.1: Point cloud reconstructed from the extracted feature points of [PITH_FULL_IMAGE:figures/full_fig_p082_5_1.png] view at source ↗
Figure 5.2
Figure 5.2. Figure 5.2: The rays emitted from an object placed at an axial distance [PITH_FULL_IMAGE:figures/full_fig_p087_5_2.png] view at source ↗
Figure 5.3
Figure 5.3. Figure 5.3: The overview of our proposed network, FocDepthFormer, is pre [PITH_FULL_IMAGE:figures/full_fig_p087_5_3.png] view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: To illustrate the LSTM module in our network, the initial step [PITH_FULL_IMAGE:figures/full_fig_p089_5_4.png] view at source ↗
Figure 5.5
Figure 5.5. Figure 5.5: Comparison of Transformer attention on the two left column [PITH_FULL_IMAGE:figures/full_fig_p090_5_5.png] view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: Specifically, for tokens within a single frame, only the activated [PITH_FULL_IMAGE:figures/full_fig_p091_5_4.png] view at source ↗
Figure 5.6
Figure 5.6. Figure 5.6: Qualitative evaluation of our model on DDFF 12-Scene dataset. [PITH_FULL_IMAGE:figures/full_fig_p096_5_6.png] view at source ↗
Figure 5.7
Figure 5.7. Figure 5.7: Qualitative evaluation of our model on Mobile Depth dataset. [PITH_FULL_IMAGE:figures/full_fig_p096_5_7.png] view at source ↗
Figure 5.8
Figure 5.8. Figure 5.8: Qualitative evaluation of our model on LightField4D dataset. [PITH_FULL_IMAGE:figures/full_fig_p096_5_8.png] view at source ↗
Figure 5.9
Figure 5.9. Figure 5.9: The top row is the input, and the bottom is the output disparity [PITH_FULL_IMAGE:figures/full_fig_p099_5_9.png] view at source ↗
Figure 5.10
Figure 5.10. Figure 5.10: Different model structures’ comparison. former + LSTM design proposed in the paper can predict a more detailed feature map with fine-grained details, while the naive concatenation of feature maps after the Transformer can not achieve the equivalent performance (as reflected in the middle image) compared to our proposed structure, by com￾bining the focus/defocus cues directly. The CNN encode even with th… view at source ↗
Figure 5.11
Figure 5.11. Figure 5.11: Our model performance w.r.t. the frame size of one focal stack [PITH_FULL_IMAGE:figures/full_fig_p102_5_11.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: 2D cross-section of the implicit SDF volume space, where red [PITH_FULL_IMAGE:figures/full_fig_p106_6_1.png] view at source ↗
Figure 6.2
Figure 6.2. Figure 6.2: Sampling points near the surface are generated to compute Signed [PITH_FULL_IMAGE:figures/full_fig_p107_6_2.png] view at source ↗
Figure 6.3
Figure 6.3. Figure 6.3: GT SDF value generation can be achieved using two main [PITH_FULL_IMAGE:figures/full_fig_p108_6_3.png] view at source ↗
Figure 6.4
Figure 6.4. Figure 6.4: 3D CNN-based UNet for shape point cloud completion, image [PITH_FULL_IMAGE:figures/full_fig_p111_6_4.png] view at source ↗
Figure 6.5
Figure 6.5. Figure 6.5: The implicit SDF model for shape reconstruction in the training [PITH_FULL_IMAGE:figures/full_fig_p112_6_5.png] view at source ↗
Figure 6.6
Figure 6.6. Figure 6.6: The implicit SDF model for shape reconstruction in the test [PITH_FULL_IMAGE:figures/full_fig_p112_6_6.png] view at source ↗
Figure 6.7
Figure 6.7. Figure 6.7: Our model is based on implicit triplane feature fusion for Signed [PITH_FULL_IMAGE:figures/full_fig_p115_6_7.png] view at source ↗
Figure 6.8
Figure 6.8. Figure 6.8: A three-stage preprocessing pipeline for distractor removal: (a) [PITH_FULL_IMAGE:figures/full_fig_p117_6_8.png] view at source ↗
Figure 6.9
Figure 6.9. Figure 6.9: (a) Raw image with distractor on the ground. (b) Inpainted image [PITH_FULL_IMAGE:figures/full_fig_p117_6_9.png] view at source ↗
Figure 6.10
Figure 6.10. Figure 6.10: Wavelet transform of the depth map in finest resolution, (a) is [PITH_FULL_IMAGE:figures/full_fig_p119_6_10.png] view at source ↗
Figure 6.11
Figure 6.11. Figure 6.11: Sampling points along the pixel ray r(t) starting from o for implicit triplane feature learning via projection. For Wavelet feature projec￾tion onto triplane. The ray r(t) starts from the camera origin o, then passes through a single unprojected point S. Dashed lines represent the orthogonal projections onto the xy, xz, and yz planes to obtain triplane features sxy, sxz, and syz for a 3D point. Where {Z… view at source ↗
Figure 6
Figure 6. Figure 6: illustrates the feature extraction process along a pixel ray [PITH_FULL_IMAGE:figures/full_fig_p121_6.png] view at source ↗
Figure 6.12
Figure 6.12. Figure 6.12: The results of 3D UNet model for point cloud completion, where [PITH_FULL_IMAGE:figures/full_fig_p124_6_12.png] view at source ↗
Figure 6.13
Figure 6.13. Figure 6.13: The shape reconstruction results from 3D voxel input via im [PITH_FULL_IMAGE:figures/full_fig_p125_6_13.png] view at source ↗
Figure 6.14
Figure 6.14. Figure 6.14: Baseline comparison results on five targets from DTU ([103]), [PITH_FULL_IMAGE:figures/full_fig_p129_6_14.png] view at source ↗
Figure 6.15
Figure 6.15. Figure 6.15: Ablation study on 3D reconstruction. From left to right: (1) Re [PITH_FULL_IMAGE:figures/full_fig_p130_6_15.png] view at source ↗
Figure 6.16
Figure 6.16. Figure 6.16: Visualization of learned wavelet encoder feature maps at the [PITH_FULL_IMAGE:figures/full_fig_p132_6_16.png] view at source ↗
Figure 6.17
Figure 6.17. Figure 6.17: Ablation study visualizations of different design combinations [PITH_FULL_IMAGE:figures/full_fig_p134_6_17.png] view at source ↗
Figure 8
Figure 8. Figure 8: illustrates pose estimation from camera images based on track [PITH_FULL_IMAGE:figures/full_fig_p144_8.png] view at source ↗
Figure 8.1
Figure 8.1. Figure 8.1: Camera pose is calculated from the live-streaming images, demo [PITH_FULL_IMAGE:figures/full_fig_p145_8_1.png] view at source ↗
Figure 8.2
Figure 8.2. Figure 8.2: Experimental Hardware Test Setup: Our camera pose estima [PITH_FULL_IMAGE:figures/full_fig_p145_8_2.png] view at source ↗
Figure 8.3
Figure 8.3. Figure 8.3: Camera pose estimation system tested both on the ground and [PITH_FULL_IMAGE:figures/full_fig_p146_8_3.png] view at source ↗
Figure 8
Figure 8. Figure 8: shows the registration of head scans in the body scan of terracotta [PITH_FULL_IMAGE:figures/full_fig_p148_8.png] view at source ↗
Figure 8.4
Figure 8.4. Figure 8.4: Registration of source and target scans of Terracotta Warriors [PITH_FULL_IMAGE:figures/full_fig_p148_8_4.png] view at source ↗
Figure 8.5
Figure 8.5. Figure 8.5: Depth estimation for the planar sculpture of Marigold. The [PITH_FULL_IMAGE:figures/full_fig_p150_8_5.png] view at source ↗
Figure 8.6
Figure 8.6. Figure 8.6: Depth prediction from focal stack technique applied to heritage [PITH_FULL_IMAGE:figures/full_fig_p151_8_6.png] view at source ↗
Figure 8.7
Figure 8.7. Figure 8.7: 3D AR Guides overlaid with the real temple. The image was [PITH_FULL_IMAGE:figures/full_fig_p154_8_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: demonstrates the beautiful interior of a historic cathedral with [PITH_FULL_IMAGE:figures/full_fig_p154_8.png] view at source ↗
Figure 8.8
Figure 8.8. Figure 8.8: The demo image is from public website link. [PITH_FULL_IMAGE:figures/full_fig_p155_8_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: exhibits another critical application of how 3D digital recon [PITH_FULL_IMAGE:figures/full_fig_p155_8.png] view at source ↗
Figure 8.9
Figure 8.9. Figure 8.9: Chinese 3A Video Game, Black-myth Wukong, with many cul [PITH_FULL_IMAGE:figures/full_fig_p156_8_9.png] view at source ↗
Figure 8.11
Figure 8.11. Figure 8.11: This can prove the great usage potential of our reconstruction [PITH_FULL_IMAGE:figures/full_fig_p156_8_11.png] view at source ↗
Figure 8.10
Figure 8.10. Figure 8.10: 3D printed copy of Dunhuang Mogao Grottes inside the moun [PITH_FULL_IMAGE:figures/full_fig_p157_8_10.png] view at source ↗
Figure 8.11
Figure 8.11. Figure 8.11: Practical application of my 3D reconstruction model: The her [PITH_FULL_IMAGE:figures/full_fig_p158_8_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

288 extracted references · 60 canonical work pages · 7 internal anchors

  1. [1]

    Satr: Zero-shot semantic segmentation of 3d shapes

    Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., and Wonka, P. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2023), pp. 15166–15179

    2023

  2. [2]

    Slic superpixels compared to state-of-the-art super- pixel methods

    Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., and Susstrunk, S. Slic superpixels compared to state-of-the-art super- pixel methods. IEEE transactions on pattern analysis and machine intelligence 34 (05 2012)

    2012

  3. [3]

    Depthformer: Multiscale vision transformer for monocular depth estimation with global local infor- mation fusion

    Agarwal, A., and Arora, C. Depthformer: Multiscale vision transformer for monocular depth estimation with global local infor- mation fusion. In 2022 IEEE International Conference on Image Pro- cessing (ICIP) (2022), IEEE, pp. 3873–3877

    2022

  4. [4]

    Alpert, B. K. Hybrid gauss-trapezoidal quadrature rules. SIAM Journal on Scientific Computing 20 , 5 (1999), 1551–1584

    1999

  5. [5]

    Video sta- bilization of the nao robot using imu data

    Alquisiris-Quecha, O., and Martinez-Carranza, J. Video sta- bilization of the nao robot using imu data. In Robot Operating System (ROS). Springer, Cham, 2020, pp. 147–162

    2020

  6. [6]

    Yolopose: Transformer-based multi-object 6d pose estimation using keypoint re- gression

    Amini, A., Selvam Periyasamy, A., and Behnke, S. Yolopose: Transformer-based multi-object 6d pose estimation using keypoint re- gression. In International Conference on Intelligent Autonomous Sys- tems (2022), Springer, pp. 392–406

    2022

  7. [7]

    Deblur and deep depth from single defocus image

    Anwar, S., Hayder, Z., and Porikli, F. Deblur and deep depth from single defocus image. Machine vision and applications 32 , 1 (2021), 1–13

    2021

  8. [8]

    Spinnet: Learning a general surface descriptor for 3d point cloud registration

    Ao, S., Hu, Q., Yang, B., Markham, A., and Guo, Y. Spinnet: Learning a general surface descriptor for 3d point cloud registration. 146 In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021), pp. 11753–11762

    2021

  9. [9]

    A hybrid motion esti- mation for video stabilization based on an imu sensor

    Auysakul, J., Xu, H., and Pooneeth, V. A hybrid motion esti- mation for video stabilization based on an imu sensor. Sensors 18 , 8 (2018), 2708

    2018

  10. [10]

    Multi-view depth estima- tion by fusing single-view depth probability with multi-view geometry

    Bae, G., Budvytis, I., and Cipolla, R. Multi-view depth estima- tion by fusing single-view depth probability with multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 2842–2851

    2022

  11. [11]

    Pointdsc: Robust point cloud registration using deep spatial consistency

    Bai, X., Luo, Z., Zhou, L., Chen, H., Li, L., Hu, Z., Fu, H., and Tai, C.-L. Pointdsc: Robust point cloud registration using deep spatial consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021), pp. 15859–15869

    2021

  12. [12]

    D3feat: Joint learning of dense detection and description of 3d local features

    Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., and Tai, C.-L. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), pp. 6359–6367

    2020

  13. [13]

    Extracting the depth and all-in-focus image from a focal stack

    Barratt, S., and Hannel, B. Extracting the depth and all-in-focus image from a focal stack. In Proceedings of the IEEE International Conference on Computer Vision (2015), pp. 3451–3459

    2015

  14. [14]

    Barron, J. T. A general and adaptive robust loss function. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019), pp. 4331–4339

    2019

  15. [15]

    T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P

    Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., and Srinivasan, P. P. Mip-nerf: A mul- tiscale representation for anti-aliasing neural radiance fields. In Pro- ceedings of the IEEE/CVF international conference on computer vision (2021), pp. 5855–5864

    2021

  16. [16]

    T., Mildenhall, B., Verbin, D., Srinivasan, P

    Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. Mip-nerf 360: Unbounded anti-aliased neural radi- ance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022), pp. 5470–5479

    2022

  17. [17]

    Sift features tracking for video stabilization

    Battiato, S., et al. Sift features tracking for video stabilization. In 14th International Conference on Image Analysis and Processing (ICIAP 2007) (2007), IEEE. 147

    2007

  18. [18]

    Efficient surfel-based slam using 3d laser range data in urban environments

    Behley, J., and Stachniss, C. Efficient surfel-based slam using 3d laser range data in urban environments. In Robotics: Science and Systems (2018), vol. 2018, p. 59

    2018

  19. [19]

    T., Ignatov, A., and Timofte, R.Phonedepth: A dataset for monocular depth estimation on mobile devices

    Benavides, F. T., Ignatov, A., and Timofte, R.Phonedepth: A dataset for monocular depth estimation on mobile devices. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 3049–3056

    2022

  20. [20]

    G., Russo, M., Giordano, A., and Spallone, R

    Bevilacqua, M. G., Russo, M., Giordano, A., and Spallone, R. 3d reconstruction, digital twinning, and virtual reality: Architec- tural heritage applications. In 2022 IEEE Conference on Virtual Re- ality and 3D User Interfaces Abstracts and Workshops (VR W) (2022), IEEE, pp. 92–96

    2022

  21. [21]

    L., Sminchisescu, C., Theobalt, C., and Pons- Moll, G

    Bhatnagar, B. L., Sminchisescu, C., Theobalt, C., and Pons- Moll, G. Loopreg: Self-supervised learning of implicit surface corre- spondences, pose and shape for 3d human mesh registration. Advances in Neural Information Processing Systems 33 (2020), 12909–12922

    2020

  22. [22]

    Learning less is more-6d camera localization via 3d surface regression

    Brachmann, E., and Rother, C. Learning less is more-6d camera localization via 3d surface regression. In Proceedings of the IEEE con- ference on computer vision and pattern recognition (2018), pp. 4654– 4662

    2018

  23. [23]

    M., Bruna, J., LeCun, Y., Szlam, A., and V an- dergheynst, P

    Bronstein, M. M., Bruna, J., LeCun, Y., Szlam, A., and V an- dergheynst, P. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine 34 , 4 (2017), 18–42

    2017

  24. [24]

    Fourier image transformer

    Buchholz, T.-O., and Jug, F. Fourier image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 1846–1854

    2022

  25. [25]

    J., Montiel, J

    Campos, C., Elvira, R., Gomez, J. J., Montiel, J. M. M., and Tardos, J. D. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics 37, 6 (2021), 1874–1890

    2021

  26. [26]

    A computational approach to edge detection.IEEE Transactions on Pattern Analysis and Machine Intelligence 6 (1986), 679–698

    Canny, J., et al. A computational approach to edge detection.IEEE Transactions on Pattern Analysis and Machine Intelligence 6 (1986), 679–698. 148

    1986

  27. [27]

    Attitude estimation using horizon detection in thermal images

    Carrio, A., Bavle, H., and Campoy, P. Attitude estimation using horizon detection in thermal images. International Journal of Micro Air Vehicles 10 , 4 (2018), 352–361

    2018

  28. [28]

    Carvalho, M., Le Saux, B., Trouv ´e-Peloux, P., Almansa, A., and Champagnat, F. Deep depth from defocus: how can de- focus blur improve 3d estimation using dense neural networks? In Proceedings of the European Conference on Computer Vision (ECCV) Workshops (2018), pp. 0–0

    2018

  29. [29]

    Depth prediction without the sensors: Leveraging structure for unsu- pervised learning from monocular videos

    Casser, V., Pirk, S., Mahjourian, R., and Angelova, A. Depth prediction without the sensors: Leveraging structure for unsu- pervised learning from monocular videos. In Proceedings of the AAAI conference on artificial intelligence (2019), vol. 33, pp. 8001–8008

    2019

  30. [30]

    A program to build e (n)- equivariant steerable cnns

    Cesa, G., Lang, L., and Weiler, M. A program to build e (n)- equivariant steerable cnns. In International conference on learning rep- resentations (2022)

    2022

  31. [31]

    E., Ilg, E., Schmidt, T., Straub, J., Lovegrove, S., and Newcombe, R

    Chabra, R., Lenssen, J. E., Ilg, E., Schmidt, T., Straub, J., Lovegrove, S., and Newcombe, R. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16 (2020), Springer, pp. 608–625

    2020

  32. [32]

    G., Yu, B., and Vetterli, M.Adaptive wavelet thresh- olding for image denoising and compression

    Chang, S. G., Yu, B., and Vetterli, M.Adaptive wavelet thresh- olding for image denoising and compression. IEEE transactions on image processing 9, 9 (2000), 1532–1546

    2000

  33. [33]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction

    Chen, D., Li, H., Ye, W., W ang, Y., Xie, W., Zhai, S., W ang, N., Liu, H., Bao, H., and Zhang, G. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521 (2024)

    Pith/arXiv arXiv 2024

  34. [34]

    Equiv- ariant point network for 3d point cloud analysis

    Chen, H., Liu, S., Chen, W., Li, H., and Hill, R. Equiv- ariant point network for 3d point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021), pp. 14514–14523

    2021

  35. [35]

    A., and Brachmann, E

    Chen, S., Cavallari, T., Prisacariu, V. A., and Brachmann, E. Map-relative pose regression for visual re-localization. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 20665–20674. 149

    2024

  36. [36]

    Learn- ing a wavelet-like auto-encoder to accelerate deep neural networks

    Chen, T., Lin, L., Zuo, W., Luo, X., and Zhang, L. Learn- ing a wavelet-like auto-encoder to accelerate deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence (2018), vol. 32

    2018

  37. [37]

    G., and Gui, L.-Y

    Cheng, Y.-C., Lee, H.-Y., Tulyakov, S., Schwing, A. G., and Gui, L.-Y. Sdfusion: Multimodal 3d shape completion, reconstruc- tion, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 4456–4465

    2023

  38. [38]

    Implicit func- tions in feature space for 3d shape reconstruction and completion

    Chibane, J., Alldieck, T., and Pons-Moll, G. Implicit func- tions in feature space for 3d shape reconstruction and completion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (jun 2020), IEEE

    2020

  39. [39]

    Self-Supervised Real-time Video Stabilization

    Choi, J., Park, J., S, I., and Kweon . Self-supervised real-time video stabilization, 2021. arXiv preprint arXiv:2111.05980

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  40. [40]

    Diffusion-sdf: Conditional generative modeling of signed distance functions

    Chou, G., Bahat, Y., and Heide, F. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF international conference on computer vision (2023), pp. 2262–2272

    2023

  41. [41]

    Deep global registration

    Choy, C., Dong, W., and Koltun, V. Deep global registration. In CVPR (2020)

    2020

  42. [42]

    In Proceedings of the IEEE/CVF international conference on computer vision (2019), pp

    Choy, C., Park, J., and Koltun, V.Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision (2019), pp. 8958–8966

    2019

  43. [43]

    B., Xu, D., Gwak, J., Chen, K., and Savarese, S

    Choy, C. B., Xu, D., Gwak, J., Chen, K., and Savarese, S. 3d-r2n2: A unified approach for single and multi-view 3d object re- construction. In Proceedings of the European Conference on Computer Vision (ECCV) (2016)

    2016

  44. [44]

    Real-time camera pose estimation for sports fields

    Citraro, L., M ´arquez-Neila, P., Savare, S., Jayaram, V., Dubout, C., Renaut, F., Hasfura, A., Ben Shitrit, H., and Fua, P. Real-time camera pose estimation for sports fields. Machine Vision and Applications 31 , 3 (2020), 16

    2020

  45. [45]

    S., Geiger, M., K¨ohler, J., and Welling, M

    Cohen, T. S., Geiger, M., K¨ohler, J., and Welling, M. Spher- ical cnns. arXiv preprint arXiv:1801.10130 (2018). 150

    Pith/arXiv arXiv 2018

  46. [46]

    Surfel based geometry reconstruction

    Dahl, V., Aanæs, H., and Bærentzen, J. Surfel based geometry reconstruction. pp. 39–44

  47. [47]

    Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasac- chi, A., and Guibas, L. J. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (2021), pp. 12200–12209

    2021

  48. [48]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  49. [49]

    Dh3d: Deep hierarchical 3d descriptors for robust large-scale 6dof relocalization

    Du, J., W ang, R., and Cremers, D. Dh3d: Deep hierarchical 3d descriptors for robust large-scale 6dof relocalization. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16 (2020), Springer, pp. 744–762

    2020

  50. [50]

    Se(3) equivariant graph neural networks with complete local frames

    Du, W., Zhang, H., Du, Y., Meng, Q., Chen, W., Zheng, N., Shao, B., and Liu, T.-Y. Se(3) equivariant graph neural networks with complete local frames. In International Conference on Machine Learning (2022), PMLR, pp. 5583–5608

    2022

  51. [51]

    Fixed-wing attitude estimation using computer vision based horizon detection

    Dusha, D., Boles, W., and W alker, R. Fixed-wing attitude estimation using computer vision based horizon detection. In Proceed- ings of AIAC12: 2nd Australasian Unmanned Air Vehicles Conference (2007), Waldron Smith Management

    2007

  52. [52]

    Depth map prediction from a single image using a multi-scale deep network

    Eigen, D., Puhrsch, C., and Fergus, R. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)

    2014

  53. [53]

    J.A point set generation network for 3d object reconstruction from a single image

    F an, H., Su, H., and Guibas, L. J.A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition (2017), pp. 605–613

    2017

  54. [54]

    Unsupervised Monocular Depth Prediction for Indoor Continuous Video Streams

    Feng, Y., Wu, S., K ¨op¨ukl¨u, O., Kang, X., and Tombari, F. Unsupervised monocular depth prediction for indoor continuous video streams. arXiv preprint arXiv:1911.08995 (2019). 151

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  55. [55]

    A., and Nowak, R

    Figueiredo, M. A., and Nowak, R. D. An em algorithm for wavelet-based image restoration. IEEE Transactions on Image Pro- cessing 12, 8 (2003), 906–916

    2003

  56. [56]

    Finzi, M., Stanton, S., Izmailov, P., and Wilson, A. G. Gen- eralizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning (2020), PMLR, pp. 3165–3176

    2020

  57. [57]

    A., and Bolles, R

    Fischler, M. A., and Bolles, R. C. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24 , 6 (1981), 381–395

    1981

  58. [58]

    Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image

    Fu, X., Yin, W., Hu, M., W ang, K., Ma, Y., Tan, P., Shen, S., Lin, D., and Long, X. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In ECCV (2024)

    2024

  59. [59]

    Se (3)-transformers: 3d roto-translation equivariant attention networks

    Fuchs, F., Worrall, D., Fischer, V., and Welling, M. Se (3)-transformers: 3d roto-translation equivariant attention networks. Advances in neural information processing systems 33 (2020), 1970– 1981

    2020

  60. [60]

    B., W agstaff, E., Dauparas, J., and Posner, I

    Fuchs, F. B., W agstaff, E., Dauparas, J., and Posner, I. Iterative se (3)-transformers. In Geometric Science of Information: 5th International Conference, GSI 2021, Paris, France, July 21–23, 2021, Proceedings 5 (2021), Springer, pp. 585–595

    2021

  61. [61]

    Fuentes-Pacheco, J., Ruiz-Ascencio, J., and Rend ´on- Mancha, J. M. Visual simultaneous localization and mapping: a survey. Artificial intelligence review 43 (2015), 55–81

    2015

  62. [62]

    Wavelet Convolutional Neural Networks

    Fujieda, S., Takayama, K., and Hachisuka, T. Wavelet convo- lutional neural networks. arXiv preprint arXiv:1805.08620 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  63. [63]

    Deep depth from focal stack with defocus model for camera-setting invariance

    Fujimura, Y., Iiyama, M., Funatomi, T., and Mukaigawa, Y. Deep depth from focal stack with defocus model for camera-setting invariance. International Journal of Computer Vision (2023), 1–16

    2023

  64. [64]

    In 2013 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (2013), IEEE, pp

    Furgale, P., Rehder, J., and Siegwart, R.Unified temporal and spatial calibration for multi-sensor systems. In 2013 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (2013), IEEE, pp. 1280–1286. 152

    2013

  65. [65]

    K., Carneiro, G., and Reid, I

    Garg, R., Bg, V. K., Carneiro, G., and Reid, I. Unsuper- vised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision–ECCV 2016: 14th European Conference, Amster- dam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14 (2016), Springer, pp. 740–756

    2016

  66. [66]

    Are we ready for au- tonomous driving? the kitti vision benchmark suite

    Geiger, A., Lenz, P., and Urtasun, R. Are we ready for au- tonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition (2012), IEEE, pp. 3354–3361

    2012

  67. [67]

    Real- time rgb-d camera relocalization

    Glocker, B., Izadi, S., Shotton, J., and Criminisi, A. Real- time rgb-d camera relocalization. In 2013 IEEE International Sym- posium on Mixed and Augmented Reality (ISMAR) (2013), IEEE, pp. 173–179

    2013

  68. [68]

    J.Unsupervised monocular depth estimation with left-right consistency

    Godard, C., Mac Aodha, O., and Brostow, G. J.Unsupervised monocular depth estimation with left-right consistency. In Proceed- ings of the IEEE conference on computer vision and pattern recognition (2017), pp. 270–279

    2017

  69. [69]

    Godard, C., Mac Aodha, O., Firman, M., and Brostow, G. J. Digging into self-supervised monocular depth estimation. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision (2019), pp. 3828–3838

    2019

  70. [70]

    Godard, C., Mac Aodha, O., Firman, M., and Brostow, G. J. Digging into self-supervised monocular depth prediction

  71. [71]

    D., Guibas, L

    Gojcic, Z., Zhou, C., Wegner, J. D., Guibas, L. J., and Birdal, T. Learning multiview 3d point cloud registration. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), pp. 1759–1769

    2020

  72. [72]

    Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras

    Gordon, A., Li, H., Jonschkowski, R., and Angelova, A. Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision (2019), pp. 8977–8986

    2019

  73. [73]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering

    Gu´edon, A., and Lepetit, V. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 5354–5363. 153

    2024

  74. [74]

    In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (2017), pp

    Guo, T., Seyed Mousavi, H., Huu Vu, T., and Monga, V.Deep wavelet prediction for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (2017), pp. 104–113

    2017

  75. [75]

    Learning monoc- ular depth by distilling cross-domain stereo networks

    Guo, X., Li, H., Yi, S., Ren, J., and W ang, X. Learning monoc- ular depth by distilling cross-domain stereo networks. In Proceed- ings of the European Conference on Computer Vision (ECCV) (2018), pp. 484–500

    2018

  76. [76]

    Single image depth estimation trained via depth from defocus cues

    Gur, S., and Wolf, L. Single image depth estimation trained via depth from defocus cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019), pp. 7683–7692

    2019

  77. [77]

    Particle filter theory and practice with posi- tioning applications

    Gustafsson, F., et al. Particle filter theory and practice with posi- tioning applications. IEEE Aerospace and Electronic Systems Magazine 25, 7 (2010), 53–82

    2010

  78. [78]

    Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction

    Hasson, Y., Tekin, B., Bogo, F., Laptev, I., Pollefeys, M., and Schmid, C. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), pp. 571–580

    2020

  79. [79]

    G., Staab, M

    Hazirbas, C., Soyer, S. G., Staab, M. C., Leal-Taix ´e, L., and Cremers, D. Deep depth from focus. In Asian conference on computer vision (2018), Springer, pp. 525–541

    2018

  80. [80]

    Lotus: Diffusion-based visual foundation model for high-quality dense prediction

    He, J., Li, H., Yin, W., Liang, Y., Li, L., Zhou, K., Liu, H., Liu, B., and Chen, Y.-C. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124 (2024)

    Pith/arXiv arXiv 2024

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.