Pith. sign in

REVIEW 2 major objections 9 minor 107 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

6M-parameter depth model runs real-time on phones, nears foundation-model accuracy

2026-07-10 01:27 UTC pith:ZNPBJ5MS

load-bearing objection Lightweight zero-shot depth via distillation — solid engineering, baseline fairness is the real question the 2 major comments →

arxiv 2607.08771 v1 pith:ZNPBJ5MS submitted 2026-07-09 cs.CV

ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device

Fabio Tosi , Luca Bartolomei , Matteo Poggi , Stefano Mattoccia This is my paper
classification cs.CV
keywords depthfoundationlightweightmodelsmonocularzero-shotzipdepthaccuracy
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 paper argues that zero-shot depth generalization—the ability to estimate depth in scenes never seen during training—need not be the exclusive province of massive foundation models. By distilling knowledge from a large teacher model into a 6.1M-parameter convolutional network trained on 14.1M multi-domain images, the authors show that a compact architecture can retain most of the teacher's cross-domain accuracy while running at real-time rates on hardware ranging from server GPUs to 5-watt smartphones. The central mechanism is a two-part design: a reparameterizable encoder-decoder that fuses multi-branch training-time structures into single convolutions at inference, and a hardware-adaptive upsampling module that selects between a GPU-optimized convex-combination path and a mobile-friendly interpolation path before training. The paper's core claim is that the combination of broad multi-domain supervision and architecture-level efficiency design yields a model occupying a previously empty point on the accuracy-efficiency frontier: far more generalizable than prior lightweight models (which were trained on single domains), and far cheaper to deploy than foundation models (which exceed its parameter count by 50× or more).

Core claim

The paper demonstrates that the accuracy gap between a 6.1M-parameter student network and a 335M-parameter foundation-model teacher can be narrowed to roughly 1.6–4.6 percentage points of absolute relative error across five diverse benchmarks, while consuming 137× less energy per frame on embedded hardware. The key empirical finding is that this convergence depends on both ingredients simultaneously: the same architecture trained on 1% of the data degrades by 3–4 points on indoor benchmarks, and a larger data scale without the architectural skip-path and upsampling design loses fine-grained depth boundaries. The two factors are not additive but interdependent—the compact architecture must be

What carries the argument

The reparameterizable convolutional block (RepVGG-style), which maintains three parallel branches during training (3×3 conv, 1×1 conv, identity) and algebraically fuses them into a single 3×3 convolution at inference; the split-stem half-resolution skip connection that feeds boundary cues directly to the decoder; and the hardware-adaptive convex upsampling that offers two paths—a GPU path using unfold/softmax/PixelShuffle for sharp boundary interpolation, and an NPU path using a learned gate to blend nearest-neighbor and bilinear upsampling. Training uses scale-and-shift-invariant loss with multi-scale gradient loss, distilling pseudo-labels from Depth Anything v2-Large across 17 image-domai

Load-bearing premise

The claim that ZipDepth achieves the best accuracy-efficiency trade-off among lightweight models depends on the fairness of retraining baselines (Lite-Mono, GuideDepth, FastDepth, PyDNet) on the authors' multi-domain dataset. When exact protocol replication was infeasible due to architectural constraints, the authors approximated by matching batch size and iteration count, which may under-optimize baselines and make the trade-off ranking less reliable.

What would settle it

A retrained baseline (e.g., Lite-Mono) given a better-tuned training schedule, longer convergence, or architecture-specific data augmentation that matches its inductive biases could match or exceed ZipDepth's zero-shot accuracy at comparable parameter count and energy budget, breaking the claimed trade-off frontier.

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

If this is right

  • If a 6M-parameter model can approximate a 335M-parameter teacher within a few error points, the practical deployment of zero-shot depth estimation on drones, mobile AR, and edge SLAM systems becomes feasible without hardware compromise.
  • The distillation recipe—broad multi-domain data plus a teacher providing pseudo-labels—may transfer to other dense prediction tasks (normal estimation, segmentation, optical flow) where foundation models exist but are too heavy for edge deployment.
  • The hardware-adaptive upsampling split suggests a design pattern where a single trained model ships with two inference graphs, each optimized for a different runtime backend, reducing the engineering burden of multi-platform deployment.
  • The data-scaling ablation showing continued accuracy improvement up to 14.1M images implies that the compact architecture has not saturated its capacity to absorb supervision, leaving open whether further data would continue to close the gap to the teacher.

Where Pith is reading between the lines

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

  • If the teacher's pseudo-labels are systematically biased in certain domains (e.g., transparent surfaces, extreme lighting), the student will inherit those biases at lower resolution, potentially making the compact model less reliable than its benchmark numbers suggest in deployment scenarios outside the evaluation distribution.
  • The 50× parameter reduction with only a ~3-point accuracy gap raises the question of whether foundation models are over-parameterized for the depth-estimation task specifically, or whether the student is approaching an information-theoretic ceiling for what can be distilled from monocular RGB input.
  • The temporal flickering limitation acknowledged by the authors suggests that a lightweight recurrent or flow-based module added to the decoder could yield temporally consistent video depth at still-modest parameter counts, since the per-frame representation is already efficient.

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

2 major / 9 minor

Summary. The manuscript presents ZipDepth, a 6.1M-parameter monocular depth estimation network that combines a reparameterizable convolutional encoder-decoder with large-scale knowledge distillation from Depth Anything v2-Large over 14.1M multi-domain images. The central claim is that ZipDepth achieves the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models, running at real-time rates across nine hardware platforms spanning server GPUs to 5W smartphones. The architecture features a split-stem half-resolution skip connection, multi-scale context modules, strip-pooling attention, and a hardware-adaptive convex upsampling with dual paths (GPU/TensorRT and NPU/mobile). Evaluation covers five zero-shot benchmarks (NYUv2, KITTI, ETH3D, ScanNet, DIODE) and includes ablations for each architectural component, data scale, and upsampling strategy.

Significance. The paper addresses a genuine and important gap in the monocular depth estimation literature: the dichotomy between accurate but heavy foundation models and efficient but domain-bound lightweight models. The approach of distilling a foundation model's zero-shot capability into a compact CNN is well-motivated, and the engineering is thorough — the hardware-adaptive upsampling design with two deployment paths, the reparameterizable blocks for inference-time fusion, and the broad deployment profiling across nine platforms are concrete strengths. The data-scale ablation (Table 4) provides falsifiable evidence that the compact architecture benefits from large-scale supervision rather than saturating early. The training cost (two RTX 3090 GPUs over three days) is notably accessible for the result quality. The evaluation protocol follows established methodology (Marigold's benchmark), and the decision to retrain lightweight baselines on the same multi-domain data, rather than comparing against single-domain self-supervised weights, is methodologically sound in principle.

major comments (2)
  1. §4.1 (Baselines paragraph): The claim that ZipDepth achieves the best accuracy–efficiency trade-off among lightweight models rests on Table 2, where baselines are retrained on the authors' multi-domain dataset. The paper states that 'when exact replication is not feasible due to architectural constraints, we approximate the protocol by matching the batch size and total number of training iterations.' This is load-bearing because the trade-off is most vulnerable for efficiency-comparable baselines: PyDNet (34.8 FPS, 398.0 mJ vs. ZipDepth's 34.4 FPS, 396.6 mJ) and FastDepth (28.9 FPS, 483.2 mJ). For PyDNet, the accuracy gap on DIODE is only 0.8 AR (23.4 vs. 22.6) and on NYUv2 is 3.0 AR (11.4 vs. 8.4). The paper does not specify which baselines received the 'approximate' protocol versus the exact replication, nor how the approximations differed (e.g., different learning rate schedules, loss
  2. weight configurations, or augmentation settings). Without this information, it is impossible to assess whether PyDNet and FastDepth were trained under hyperparameters suited to their architectures or simply inherited ZipDepth's schedule, which may be suboptimal for a 1.9M-parameter network originally designed for self-supervised training. Please specify per-baseline training details, or at minimum confirm that the efficiency-critical baselines (PyDNet, FastDepth) received the full protocol rather than the approximation.
minor comments (9)
  1. §3.2, Eq. (8): The softmax temperature τ is listed as a free parameter in the axiom ledger but its value is never stated in the manuscript. Please specify the value used (or state that τ=1).
  2. Table 2: The footnote marker '†' is defined as 'retrained from scratch on the same training data,' but MiDaS-Small appears without the dagger and is presumably evaluated with its original pretrained weights. Please clarify whether MiDaS-Small is zero-shot pretrained or retrained, and verify consistency.
  3. §3.1 (Multi-scale context, Stage 2): Eq. (2) uses depthwise convolutions with dilation rates r=1 and r=2, but the text mentions a '5×5 effective receptive field.' Please clarify whether this refers to the union of the two dilation fields or sequential application.
  4. Figure 3: The architecture diagram is dense and some labels (e.g., channel dimensions at fusion points, the SPPF bottleneck projection) are difficult to read. A higher-resolution or simplified version would aid review.
  5. §4.3 (Architectural components ablation, Table 3): The 'w/o Half-Res Path' variant is the fastest (48.9 FPS) but the text does not clarify whether this variant also removes the convex upsampling or replaces it with bilinear. Please specify.
  6. Table 5: The CPU column reports 47.4 FPS for bilinear and 25.6 FPS for the convex GPU path, but the text in §4.3 states 'on CPU, the unfold-based GPU path is substantially slower (25.6 vs. 44.6 FPS).' The 44.6 FPS figure appears in the table but not in the text; please verify consistency.
  7. §4.4, Table 6: The 'Base' and 'Optimized' column headers are defined in the text but the specific backend for each cell is only indicated by footnote symbols. A brief inline note (e.g., 'PyTorch FP32' vs. 'TensorRT FP16') for each row would improve readability.
  8. The abstract states '50× more parameters' while the text (§1) states '55× fewer parameters' relative to DA-V2-Large (335M / 6.1M ≈ 54.9). Please align the rounding.
  9. References [58] and [59] appear to cite NTIRE 2026 and 2025 challenge reports; please verify these are published or forthcoming and that the citation details are complete.

Circularity Check

0 steps flagged

No circularity: central claims are evaluated against external benchmarks; no fitted constants are presented as predictions; no self-citation chain is load-bearing.

full rationale

The paper's central claim — that ZipDepth achieves the best accuracy–efficiency trade-off among lightweight models — is evaluated against five external benchmarks (NYUv2, KITTI, ETH3D, ScanNet, DIODE) that are explicitly stated as unseen during training (§4.1: 'We test on five real-world datasets unseen during training'). The teacher model (Depth Anything v2-Large) is an external pretrained model from the same broader research community, not a prior work by these authors. The SSI loss and gradient loss (Eq. 11) are adopted from MiDaS [61] and Depth Anything v2 [94], both external works, and are standard formulations — no fitted constants are presented as novel predictions. The architectural components (reparameterizable blocks, SE attention, GC block, convex upsampling) are drawn from external prior work (RepVGG [15], SE [32], GCNet [9], RAFT [74]) and are not claimed as derivations from first principles. The ablation studies (Tables 3–5) remove components and measure accuracy changes on external benchmarks, which is standard empirical validation, not circular reasoning. The data-scale ablation (Table 4) varies training data while holding architecture fixed, showing monotonic improvement — an independent empirical result. No equation in the paper reduces to its own input by construction. The only mild concern is the fairness of baseline retraining (§4.1), but this is a correctness/experimental-design concern, not a circularity issue — the baselines are evaluated on the same external benchmarks. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 2 invented entities

The free parameters are standard hyperparameters for depth network training, mostly inherited from the teacher's configuration. The key axioms are domain assumptions about teacher quality and benchmark representativeness. The two invented entities (dual-path upsampling, split-stem skip) are both ablated with independent evidence.

free parameters (9)
  • λ_ssi = 1.0
    Loss weight for SSI term, set to match Depth Anything v2 configuration.
  • λ_grad = 2.0
    Loss weight for multi-scale gradient term, set to match Depth Anything v2 configuration.
  • C_1 (stem channels) = 48
    Channel width for stem convolution, chosen by design.
  • C_half (half-res skip channels) = 32
    Channel width for half-resolution skip connection.
  • τ (softmax temperature for convex upsampling) = not specified
    Temperature parameter in the convex upsampling softmax normalization (Eq. 8), value not given in the main text.
  • SE reduction ratio = 4
    Reduction ratio for Squeeze-and-Excitation bottleneck in Stage 3.
  • GC block reduction ratio = 4
    Reduction ratio for Global Context block bottleneck.
  • Training resolutions = 256→384→512
    Progressive resolution schedule, chosen empirically.
  • Peak learning rates per stage = 2e-3, 5e-4, 2.5e-4
    Learning rates for each progressive training stage, chosen empirically.
axioms (4)
  • domain assumption Depth Anything v2-Large produces sufficiently accurate pseudo-labels across all 17 training domains to serve as a reliable teacher.
    The entire training pipeline depends on teacher quality; if the teacher fails on certain domains (e.g., transparent surfaces, adverse weather), the student inherits those failures. Invoked in §1 and §4.1.
  • domain assumption Affine-invariant (SSI) depth prediction is a useful output format for downstream applications.
    The model predicts relative depth up to scale and shift; applications requiring metric depth need additional calibration. Invoked in §3.3.
  • domain assumption The five evaluation benchmarks (NYUv2, KITTI, ETH3D, ScanNet, DIODE) are representative of real-world deployment domains.
    Zero-shot generalization is claimed based on performance on these five datasets; domains not represented (e.g., underwater, aerial) are only qualitatively assessed. Invoked in §4.1.
  • standard math Reparameterizable multi-branch blocks trained with BN can be losslessly fused to single convolutions at inference.
    This is a known algebraic identity from RepVGG [15]; the paper correctly applies it. Invoked in §3.1.
invented entities (2)
  • Hardware-adaptive convex upsampling (dual-path) independent evidence
    purpose: Provides a GPU-optimized path (unfold + softmax + PixelShuffle) and a mobile-NPU-friendly path (learned gate blending nearest and bilinear), selected before training based on target hardware.
    Both paths are evaluated in Table 5 with FPS and boundary F1 metrics on different hardware, showing the NPU path is 1.56× faster on iPhone ANE while the GPU path yields sharper boundaries (SI-BF1 0.120 vs 0.105).
  • Split-stem half-resolution skip connection independent evidence
    purpose: Preserves boundary cues at H/2 resolution for the decoder without maintaining full-resolution feature maps in the encoder.
    Ablated in Table 3: removing it degrades accuracy on all five benchmarks (e.g., NYUv2 AbsRel 8.4→9.1) while increasing FPS (34.4→48.9), confirming its contribution to boundary preservation.

pith-pipeline@v1.1.0-glm · 23734 in / 3067 out tokens · 165159 ms · 2026-07-10T01:27:33.089737+00:00 · methodology

0 comments
read the original abstract

Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place them far beyond the reach of embedded and mobile platforms. Lightweight alternatives exist, but have been developed almost exclusively within single-domain, self-supervised paradigms, failing silently under domain shift. We present ZipDepth, a compact monocular depth network that bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model over a large multi-domain training set. Comprising just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, taking a significant step towards the accuracy of foundation models with 50x more parameters.

Figures

Figures reproduced from arXiv: 2607.08771 by Fabio Tosi, Luca Bartolomei, Matteo Poggi, Stefano Mattoccia.

Figure 1
Figure 1. Figure 1: ZipDepth generalizes zero-shot across diverse and challenging scenes. Top row: input RGB; bottom row: predicted depth at real-time rates even on a 15W Jetson Orin NX (34 FPS with PyTorch Eager FP32, up to 77 FPS with TensorRT FP16). Abstract. Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place the… view at source ↗
Figure 2
Figure 2. Figure 2: Energy–Accuracy Trade-off. Zero-shot depth models on NVIDIA Jetson Orin NX (15W). Energy per frame (FP32 inference) vs. mean AbsRel (over 5 datasets); bubble size indicates FPS. ZipDepth bridges the gap between lightweight and founda￾tion models, tending toward their accuracy under a < 400 mJ/frame budget. † denotes lightweight baselines retrained on our multi-domain training set for a fair comparison. sys… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the ZipDepth architecture. The encoder (top, left→right) progressively downsamples the input through four reparameterizable stages; the de￾coder (bottom, right→left) fuses multi-scale features coarse-to-fine and produces a full-resolution inverse-depth map via hardware-adaptive convex upsampling. 3 Method As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison. Depth maps predicted by each method on five unseen benchmarks. All lightweight models are retrained on the same data. The DA￾V2-Large teacher is shown as a large-model reference [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of upsampling strategies. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Zero-shot depth in the wild. ZipDepth on challenging DA-2K [94] scenes, generalizing far beyond its evaluation domains. No ground truth available. 5 Conclusion We presented ZipDepth, a lightweight monocular depth network bridging the zero-shot generalization of foundation models with the efficiency demands of em￾bedded deployment. Combining a reparameterizable encoder, hardware-adaptive convex upsampling, … 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

107 extracted references · 107 canonical work pages · 13 internal anchors

  1. [1]

    In: Proceedings of the European conference on computer vision (ECCV) workshops

    Aleotti, F., Tosi, F., Poggi, M., Mattoccia, S.: Generative adversarial networks for unsupervised monocular depth prediction. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0–0 (2018) 4

    work page 2018

  2. [2]

    In: European Conference on Computer Vision

    Aleotti, F., Tosi, F., Zhang, L., Poggi, M., Mattoccia, S.: Reversing the cycle: self- supervised deep stereo through enhanced monocular distillation. In: European Conference on Computer Vision. pp. 614–632. Springer (2020) 4

    work page 2020

  3. [3]

    Sensors 21(1), 15 (2020) 5

    Aleotti, F., Zaccaroni, G., Bartolomei, L., Poggi, M., Tosi, F., Mattoccia, S.: Real- time single image depth perception in the wild with handheld devices. Sensors 21(1), 15 (2020) 5

    work page 2020

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009–4018 (2021) 4

    work page 2021

  5. [5]

    In: Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part I

    Bhat, S.F., Alhashim, I., Wonka, P.: Localbins: Improving depth estimation by learning local distributions. In: Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part I. pp. 480–

    work page 2022

  6. [6]

    Springer (2022) 2, 4

    work page 2022

  7. [7]

    ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth

    Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    In: Advances in Neural Information Processing Systems

    Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsuper- vised Scale-consistent Depth and Ego-motion Learning from Monocular Video. In: Advances in Neural Information Processing Systems. vol. 32 (2019) 4

    work page 2019

  9. [9]

    Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

    Bochkovskii, A., Delaunoy, A., Germain, H., Santos, M., Zhou, Y., Richter, S.R., Koltun, V.: Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073 (2024) 5, 13, 14

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    In: Proceedings of the IEEE/CVF international conference on computer vision workshops

    Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze- excitation networks and beyond. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0–0 (2019) 7

    work page 2019

  11. [11]

    In: Proceed- ings of the Computer Vision and Pattern Recognition Conference

    Chen, S., Guo, H., Zhu, S., Zhang, F., Huang, Z., Feng, J., Kang, B.: Video depth anything: Consistent depth estimation for super-long videos. In: Proceed- ings of the Computer Vision and Pattern Recognition Conference. pp. 22831– 22840 (2025) 4

    work page 2025

  12. [12]

    Advances in neural information processing systems29(2016) 4

    Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. Advances in neural information processing systems29(2016) 4

    work page 2016

  13. [13]

    IEEE Internet of Things Journal9(1), 25–36 (2021) 5

    Cipolletta,A.,Peluso,V.,Calimera,A.,Poggi,M.,Tosi,F.,Aleotti,F.,Mattoccia, S.: Energy-quality scalable monocular depth estimation on low-power cpus. IEEE Internet of Things Journal9(1), 25–36 (2021) 5

    work page 2021

  14. [14]

    In: Proc

    Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 10

    work page 2016

  15. [15]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scan- net: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828–5839 (2017) 10

    work page 2017

  16. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733–13742 (2021) 6 ZipDepth 17

    work page 2021

  17. [17]

    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) 2

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  18. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10786–10796 (2021) 4, 11

    work page 2021

  19. [19]

    In: International Con- ference on Computer Vision

    Eigen, D., Fergus, R.: Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-scale Convolutional Architecture. In: International Con- ference on Computer Vision. pp. 2650–2658 (2015).https://doi.org/10.1109/ ICCV.2015.3042, 4

    work page arXiv 2015

  20. [20]

    Advances in neural information processing sys- tems27(2014) 10

    Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing sys- tems27(2014) 10

    work page 2014

  21. [21]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2002–2011 (2018) 2, 4

    work page 2002

  22. [22]

    In: European Conference on Computer Vision

    Fu, X., Yin, W., Hu, M., Wang, K., Ma, Y., Tan, P., Shen, S., Lin, D., Long, X.: Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In: European Conference on Computer Vision. pp. 241–258. Springer (2024) 4

    work page 2024

  23. [23]

    In: European Conference on Com- puter Vision

    Garg, R., Kumar, V., Carneiro, G., Reid, I.: Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. In: European Conference on Com- puter Vision. pp. 740–756 (2016) 2, 4

    work page 2016

  24. [24]

    International Journal of Robotics Research32(11), 1231–1237 (2013)

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. International Journal of Robotics Research32(11), 1231–1237 (2013). https://doi.org/10.1177/02783649134912973, 4, 5

    work page doi:10.1177/02783649134912973 2013

  25. [25]

    The international journal of robotics research32(11), 1231–1237 (2013) 10

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The international journal of robotics research32(11), 1231–1237 (2013) 10

    work page 2013

  26. [26]

    Conference on Computer Vision and Pattern Recognition pp

    Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised Monocular Depth Esti- mation with Left-Right Consistency. Conference on Computer Vision and Pattern Recognition pp. 6602–6611 (2017).https://doi.org/10.1109/CVPR.2017.6994

    work page doi:10.1109/cvpr.2017.6994 2017

  27. [27]

    International Conference on Computer Vision 2019-Octob, 3827–3837 (2019).https://doi.org/10.1109/ICCV.2019.00393 2, 4

    Godard, C., Aodha, O.M., Firman, M., Brostow, G.: Digging Into Self-Supervised Monocular Depth Estimation. International Conference on Computer Vision 2019-Octob, 3827–3837 (2019).https://doi.org/10.1109/ICCV.2019.00393 2, 4

    work page doi:10.1109/iccv.2019.00393 2019

  28. [28]

    In: The IEEE International Conference on Com- puter Vision (ICCV) (2019) 10

    Gruber, T., Julca-Aguilar, F., Bijelic, M., Heide, F.: Gated2depth: Real-time dense lidar from gated images. In: The IEEE International Conference on Com- puter Vision (ICCV) (2019) 10

    work page 2019

  29. [29]

    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., Chen, Y.C.: Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    IEEE transactions on pattern analysis and machine intelligence37(9), 1904–1916 (2015) 8

    He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolu- tional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence37(9), 1904–1916 (2015) 8

    work page 1904

  31. [31]

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020) 4

    work page 2020

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hou, Q., Zhang, L., Cheng, M.M., Feng, J.: Strip pooling: Rethinking spatial pool- ing for scene parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4003–4012 (2020) 7 18 F. Tosi et al

    work page 2020

  33. [33]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018) 7

    work page 2018

  34. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence46(12), 10579–10596 (2024) 2, 4

    Hu, M., Yin, W., Zhang, C., Cai, Z., Long, X., Chen, H., Wang, K., Yu, G., Shen, C., Shen, S.: Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence46(12), 10579–10596 (2024) 2, 4

    work page 2024

  35. [35]

    DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos

    Hu, W., Gao, X., Li, X., Zhao, S., Cun, X., Zhang, Y., Quan, L., Shan, Y.: Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Holopix50k: A Large-Scale In-the-wild Stereo Image Dataset

    Hua, Y., Kohli, P., Uplavikar, P., Ravi, A., Gunaseelan, S., Orozco, J., Li, E.: Holopix50k: A large-scale in-the-wild stereo image dataset. arXiv preprint arXiv:2003.11172 (2020) 10

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  37. [37]

    org / 10

    Jocher, G.: Ultralytics yolov5 (2020).https : / / doi . org / 10 . 5281 / zenodo . 3908559,https://github.com/ultralytics/yolov58

    work page 2020

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 4

    Ke, B., Narnhofer, D., Huang, S., Ke, L., Peters, T., Fragkiadaki, K., Obukhov, A., Schindler, K.: Video depth without video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 4

    work page 2025

  39. [39]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) (2024) 2, 4, 10, 11

    Ke, B., Obukhov, A., Huang, S., Metzger, N., Daudt, R.C., Schindler, K.: Repur- posing diffusion-based image generators for monocular depth estimation. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) (2024) 2, 4, 10, 11

    work page 2024

  40. [40]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023) 10

    work page 2023

  41. [41]

    IJCV (2020) 10

    Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., Duerig, T., Ferrari, V.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV (2020) 10

    work page 2020

  42. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li, W., Huang, J., Jung, H., Zhai, G., Ramirez, P.Z., Costanzino, A., Tosi, F., Poggi, M., Di Stefano, L., Weibel, J.B., et al.: Tricky 2025 housecat6d object pose estimation challenge with specular and transparent surfaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3292–3302 (2025) 4

    work page 2025

  43. [43]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from in- ternet photos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2041–2050 (2018) 4, 10

    work page 2041

  44. [44]

    In: European confer- ence on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European confer- ence on computer vision. pp. 740–755. Springer (2014) 10

    work page 2014

  45. [45]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Liu, F., Shen, C., Lin, G.: Deep convolutional neural fields for depth estimation from a single image. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5162–5170 (2015) 4

    work page 2015

  46. [46]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2025) 4 ZipDepth 19

    Martin Garcia, G., Abou Zeid, K., Schmidt, C., de Geus, D., Hermans, A., Leibe, B.: Fine-tuning image-conditional diffusion models is easier than you think. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2025) 4 ZipDepth 19

    work page 2025

  47. [47]

    In: Proceedings of the IEEE international conference on computer vision

    Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: Proceedings of the IEEE international conference on computer vision. pp. 4990–4999 (2017) 10

    work page 2017

  48. [48]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Obukhov, A., Poggi, M., Tosi, F., Arora, R.S., Spencer, J., Russel, C., Hadfield, S., Bowden, R., Wang, S., Ma, Z., et al.: The fourth monocular depth estima- tion challenge. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 6182–6195 (2025) 5

    work page 2025

  49. [49]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez,P.,Haziza,D.,Massa,F.,El-Nouby,A.,etal.:Dinov2:Learningrobust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  50. [50]

    IEEE Transactions on Circuits and Systems for Video Technology32(3), 1524–1536 (2021) 5

    Peluso,V.,Cipolletta,A.,Calimera,A.,Poggi,M.,Tosi,F.,Aleotti,F.,Mattoccia, S.: Monocular depth perception on microcontrollers for edge applications. IEEE Transactions on Circuits and Systems for Video Technology32(3), 1524–1536 (2021) 5

    work page 2021

  51. [51]

    In: 2019 Design, Automation & Test in Europe Conference & Exhibi- tion (DATE)

    Peluso, V., Cipolletta, A., Calimera, A., Poggi, M., Tosi, F., Mattoccia, S.: En- abling energy-efficient unsupervised monocular depth estimation on armv7-based platforms. In: 2019 Design, Automation & Test in Europe Conference & Exhibi- tion (DATE). pp. 1703–1708. IEEE (2019) 5

    work page 2019

  52. [52]

    UniDepthV2: Universal Monocular Metric Depth Estimation Made Simpler

    Piccinelli, L., Sakaridis, C., Yang, Y.H., Segu, M., Li, S., Abbeloos, W., Van Gool, L.: Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110 (2025) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  53. [53]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Piccinelli, L., Yang, Y.H., Sakaridis, C., Segu, M., Li, S., Van Gool, L., Yu, F.: Unidepth: Universal monocular metric depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10106– 10116 (2024) 4

    work page 2024

  54. [54]

    In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS)

    Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: Towards real-time unsupervised monocular depth estimation on cpu. In: 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5848–5854. IEEE (2018) 3, 5, 10, 11, 12

    work page 2018

  55. [56]

    IEEE Transactions on Intelligent Transportation Systems23(10), 17342–17353 (2022) 5

    Poggi,M.,Tosi,F.,Aleotti,F.,Mattoccia,S.:Real-timeself-supervisedmonocular depth estimation without gpu. IEEE Transactions on Intelligent Transportation Systems23(10), 17342–17353 (2022) 5

    work page 2022

  56. [57]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 5314–5334 (2021) 4

    Poggi, M., Tosi, F., Batsos, K., Mordohai, P., Mattoccia, S.: On the synergies between machine learning and binocular stereo for depth estimation from images: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 5314–5334 (2021) 4

    work page 2021

  57. [58]

    In: 2018 International conference on 3d vision (3DV)

    Poggi, M., Tosi, F., Mattoccia, S.: Learning monocular depth estimation with un- supervised trinocular assumptions. In: 2018 International conference on 3d vision (3DV). pp. 324–333. IEEE (2018) 4

    work page 2018

  58. [59]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ramirez, P.Z., Tosi, F., Di Stefano, L., Timofte, R., Costanzino, A., Poggi, M., Salti, S., Mattoccia, S., Min, J., Tu, J., et al.: Ntire 2026 challenge on high- resolution depth of non-lambertian surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2786–2801 (2026) 4

    work page 2026

  59. [60]

    Tosi et al

    Ramirez, P.Z., Tosi, F., Di Stefano, L., Timofte, R., Costanzino, A., Poggi, M., Salti, S., Mattoccia, S., Zhang, Z., Yang, Y., et al.: Ntire 2025 challenge on hr 20 F. Tosi et al. depth from images of specular and transparent surfaces. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 987–1001 (2025) 4

    work page 2025

  60. [61]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021) 4, 11

    work page 2021

  61. [62]

    IEEE transactions on pattern analysis and machine intelligence44(3), 1623–1637 (2020) 4, 9, 10, 11

    Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence44(3), 1623–1637 (2020) 4, 9, 10, 11

    work page 2020

  62. [63]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 12240–12249 (2019) 4

    work page 2019

  63. [64]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022) 2, 4

    work page 2022

  64. [65]

    In: 2022 Inter- national Conference on Robotics and Automation (ICRA)

    Rudolph, M., Dawoud, Y., Güldenring, R., Nalpantidis, L., Belagiannis, V.: Lightweight monocular depth estimation through guided decoding. In: 2022 Inter- national Conference on Robotics and Automation (ICRA). pp. 2344–2350. IEEE (2022) 5, 10, 11

    work page 2022

  65. [66]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10765–10775 (2021) 10

    work page 2021

  66. [67]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017) 10

    Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017) 10

    work page 2017

  67. [68]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 4

    Shao, J., Yang, Y., Zhou, H., Zhang, Y., Shen, Y., Guizilini, V., Wang, Y., Poggi, M., Liao, Y.: Learning temporally consistent video depth from video diffusion priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 4

    work page 2025

  68. [69]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430–8439 (2019) 10

    work page 2019

  69. [70]

    In: European conference on computer vision

    Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: European conference on computer vision. pp. 746–

    work page

  70. [71]

    Springer (2012) 3, 4, 10

    work page 2012

  71. [72]

    Ciagan: Conditional identity anonymization generative adversarial networks,

    Spencer, J., Bowden, R., Hadfield, S.: DeFeat-Net: General monocular depth via simultaneous unsupervised representation learning. In: Conference on Computer Vision and Pattern Recognition. pp. 14390–14401 (2020).https://doi.org/10. 1109/CVPR42600.2020.014414

    work page arXiv 2020

  72. [73]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Spencer, J., Qian, C.S., Russell, C., Hadfield, S., Graf, E., Adams, W., Schofield, A.J., Elder, J.H., Bowden, R., Cong, H., et al.: The monocular depth estimation challenge. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 623–632 (2023) 5

    work page 2023

  73. [74]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Spencer, J., Qian, C.S., Trescakova, M., Russell, C., Hadfield, S., Graf, E.W., Adams, W.J., Schofield, A.J., Elder, J., Bowden, R., et al.: The second monocular ZipDepth 21 depth estimation challenge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3064–3076 (2023) 5

    work page 2023

  74. [75]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Spencer, J., Tosi, F., Poggi, M., Arora, R.S., Russell, C., Hadfield, S., Bowden, R., Zhou, G., Li, Z., Rao, Q., et al.: The third monocular depth estimation challenge. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1–14 (2024) 5

    work page 2024

  75. [76]

    In: European conference on computer vision

    Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: European conference on computer vision. pp. 402–419. Springer (2020) 9

    work page 2020

  76. [77]

    Conference on Computer Vision and Pattern Recognition2019-June, 9791–9801 (2019).https://doi.org/10.1109/ CVPR.2019.010034

    Tosi, F., Aleotti, F., Poggi, M., Mattoccia, S.: Learning monocular depth estima- tion infusing traditional stereo knowledge. Conference on Computer Vision and Pattern Recognition2019-June, 9791–9801 (2019).https://doi.org/10.1109/ CVPR.2019.010034

    work page arXiv 2019

  77. [78]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tosi, F., Aleotti, F., Ramirez, P.Z., Poggi, M., Salti, S., Stefano, L.D., Mattoc- cia, S.: Distilled semantics for comprehensive scene understanding from videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4654–4665 (2020) 4

    work page 2020

  78. [79]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tosi, F., Liao, Y., Schmitt, C., Geiger, A.: Smd-nets: Stereo mixture density networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8942–8952 (2021) 13, 14

    work page 2021

  79. [80]

    In: European Conference on Computer Vision

    Tosi, F., Ramirez, P.Z., Poggi, M.: Diffusion models for monocular depth estima- tion: Overcoming challenging conditions. In: European Conference on Computer Vision. pp. 236–257. Springer (2024) 4

    work page 2024

  80. [81]

    DIODE: A Dense Indoor and Outdoor DEpth Dataset

    Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M.,Basart, S.,Walter,M.R., etal.: Diode: Adenseindoorand outdoor depth dataset. arXiv preprint arXiv:1908.00463 (2019) 10

    work page internal anchor Pith review Pith/arXiv arXiv 1908

Showing first 80 references.