Pith. sign in

REVIEW 3 major objections 7 minor 67 references

1B Vision Model Beats 7B Rival by Masking Boundaries

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 ↗

T0 review · glm-5.2

2026-07-07 21:48 UTC pith:7DFRGZIK

load-bearing objection Boundary-centric masked modeling is a genuinely new pretraining idea with strong empirical results, but the data-seed confound and unproven bootstrapping mechanism need referee attention. the 3 major comments →

arxiv 2607.05247 v1 pith:7DFRGZIK submitted 2026-07-06 cs.CV

Vision Pretraining for Dense Spatial Perception

Zelin Fu , Bin Tan , Changjiang Sun , Shaohui Liu , Kecheng Zheng , Yinghao Xu , Xing Zhu , Yujun Shen
show 1 more author
Nan Xue
This is my paper
classification cs.CV
keywords self-supervised learningvision transformerboundary detectionmasked modelingdepth estimationdense predictiona-contrario detectionself-distillation
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.

Modern vision foundation models are good at recognizing what is in an image but weak at recovering precise spatial structure — where one object ends and another begins, where depth jumps, where surfaces meet. This paper argues that the root cause is that boundaries (edges, contours, shape discontinuities) are treated as downstream outputs rather than as native learning signals during pretraining. The authors propose masked boundary modeling, a self-supervised method that identifies boundary-bearing tokens using the model's own online predictions, forces those tokens into the student's masked set, and supervises their reconstruction with a categorical boundary-field target rather than the usual semantic objective. The key mechanism is a bootstrapping loop: a frozen corner-point detector plus parameter-free level-line orientation from image gradients allows even an untrained network to produce usable boundary targets, which are then validated by an a-contrario statistical test (uniform orientation distribution as the null hypothesis) and refined through EMA self-distillation. Boundary prediction is reparameterized from continuous regression into per-pixel classification over discretized bins, which makes it compatible with the centering and sharpening machinery that stabilizes standard self-distillation. The result is LingBot-Vision, a 1-billion-parameter Vision Transformer that matches or surpasses models up to 7x larger on dense spatial tasks, achieving the best NYU-Depth v2 accuracy among all compared models including the 7B DINOv3, while maintaining competitive image-level recognition. The boundary-aware advantage survives distillation into smaller models and compounds with downstream data scaling, as demonstrated by the LingBot-Depth 2.0 depth completion system.

Core claim

The central discovery is that boundary structure — line segments, contours, and shape discontinuities — can serve as a primary self-supervised learning signal for vision pretraining, not merely as a downstream output. The paper shows that a self-distillation loop can bootstrap its own boundary targets from raw images without any human annotations, external edge detectors, or pretrained backbones, by exploiting the redundancy of dense boundary-field representations: many pixels vote on the same segment, so coherent geometry emerges from noisy per-pixel predictions long before the field is pixel-accurate. The categorical reparameterization of boundary fields into discrete bins is the second en

What carries the argument

Masked boundary modeling with categorical boundary fields. The teacher predicts a dense boundary field (distance, orientation, and endpoint angles per pixel relative to the nearest line segment), reparameterized as per-pixel categorical distributions over discretized bins. Boundary-bearing tokens — patches through which a predicted boundary passes — are forced into the student's masked set. Masked tokens are routed by geometry: boundary tokens receive the categorical boundary objective, all masked tokens also receive the standard semantic iBOT objective (dual supervision). Boundary targets are generated online by the EMA teacher: a frozen single-block ViT localizes corner points, the field's

Load-bearing premise

The bootstrapping loop depends on the claim that usable boundary fields emerge even from an untrained model, specifically that corner points from a frozen single-block ViT plus parameter-free level-line orientation suffice to anchor coherent line segments from noisy teacher fields. If the frozen corner detector or the level-line guidance fails to produce reliable anchors early in training, the teacher's boundary targets could be noisy or hallucinated, destabilizing the self-d

What would settle it

If the frozen corner-point detector produces unreliable anchors on a class of images (e.g., texture-heavy scenes, low-contrast indoor environments), the a-contrario validation would either reject most candidates (leaving too few boundary tokens to train on) or accept spurious segments (feeding hallucinated structure into the student). Either failure mode would manifest as training instability or degraded dense-prediction performance on scenes where corner detection is weak.

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

If this is right

  • If boundary modeling is a general pretraining principle, then other geometric structures — curvature, junctions, symmetry axes — could serve as additional native learning signals in the same self-distillation framework.
  • The 0.3B distilled student matching 7B DINOv3 on NYU-Depth v2 suggests that spatial pretraining efficiency gains are real and transferable, potentially reshaping the compute-versus-quality frontier for embodied AI perception.
  • The compounding effect (boundary advantage widens with downstream data) implies that pretraining objective design may matter more than data scale for dense spatial tasks, challenging the assumption that more data washes out initialization differences.
  • The a-contrario validation as a built-in hallucination filter for self-distillation targets could generalize to other self-supervised settings where the teacher's own predictions risk feeding noise back into the student.

Where Pith is reading between the lines

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

  • The bootstrapping loop's dependence on a frozen corner-point detector raises the question of whether the detector's quality sets a ceiling on boundary-target quality, and whether replacing it with a learned detector that co-evolves with the backbone would help or destabilize the loop.
  • The trade-off — LingBot-Vision trails DINOv3-7B on ImageNet classification by ~1.5 points while leading on dense tasks — suggests that boundary-oriented pretraining reallocates model capacity from image-level invariances to localized structure. Whether this is a fundamental trade-off or an artifact of training duration or data scale remains open.
  • If the categorical reparameterization is what makes boundary self-distillation stable, then other continuous geometric targets (surface normals, optical flow fields, depth itself) might benefit from the same discretization-plus-centering treatment in self-supervised settings.
  • The finding that boundary tokens are trackable through video by plain cosine similarity, without any temporal supervision, implies that boundary-anchored features may be naturally temporally stable — a property that could reduce the need for video pretraining in robotics perception pipelines.

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

3 major / 7 minor

Summary. This paper introduces masked boundary modeling, a self-supervised pretraining paradigm that treats boundary structure as a native learning signal rather than a downstream output. The method extends DINO/iBOT self-distillation by (1) forcing teacher-discovered boundary-bearing tokens into the student's masked set, (2) supervising those tokens with a categorical reparameterization of dense boundary fields, and (3) bootstrapping boundary targets online from the EMA teacher using a frozen corner-point detector and parameter-free a-contrario validation. The authors scale this to LingBot-Vision (1.1B parameters, ViT-g/16) and report strong frozen-feature results on dense prediction tasks (NYUv2, ADE20K, Cityscapes, VOC, DAVIS, YouTube-VOS), including best NYUv2 RMSE among compared models. Distilled ViT-L/B/S students and an application to depth completion (LingBot-Depth 2.0) further demonstrate downstream transfer.

Significance. The paper makes a conceptually clear contribution: it identifies a specific weakness of current self-distillation methods (content-agnostic masking and under-supervision at boundaries) and proposes a principled, self-contained fix. The categorical reparameterization connecting boundary fields to the a-contrario null hypothesis is elegant and yields parameter-free validation. The proof-of-concept ablations (Table 1) cleanly isolate design choices. The downstream evaluation is extensive (Tables 2-8), and the release of checkpoints and code is a strength. The data-scaling experiment (Fig. 8) showing that the pretraining advantage compounds rather than washes out is a valuable empirical finding.

major comments (3)
  1. Sec. 4.1 explicitly lists ADE20K, NYU-Depth v2, SUN-RGBD, Cityscapes, and VOC among the retrieval seeds used to curate the 161M-image pretraining corpus. These are exactly the downstream benchmarks reported in Tables 2-8. The paper's central claim is that gains come 'from the pretraining objective rather than from a data advantage' (Sec. 4.1), but retrieval from a 2B-image pool using downstream evaluation datasets as queries will populate the pretraining corpus with images visually similar to the evaluation scenes, creating a distributional alignment that DINOv2/v3 may not share. The ImageNet-1K proof of concept (Table 1) does control for data and shows objective-driven gains, but those gains are modest (NYUv2 RMSE 0.474 to 0.440) relative to the scaled margins (Table 2: NYUv2 RMSE 0.296 vs DINOv3's 0.309). The paper does not report whether DINOv2/v3's curation pipelines also included A-
  2. Sec. 3.2, Finding 1: The bootstrapping loop depends on the claim that 'usable boundary fields emerge even from an untrained model' via corner points from a frozen single-block ViT plus parameter-free level-line orientation. Fig. 3 and Appendix A demonstrate this qualitatively on two images, but no quantitative metric is provided for how reliably corner anchors produce coherent segments across a diverse image set, nor for how this reliability evolves during early training. If the frozen corner detector or level-line guidance fails to produce reliable anchors on a non-trivial fraction of images early in training, the teacher's boundary targets could be noisy, destabilizing the self-distillation loop. A quantitative ablation reporting, e.g., the fraction of training images retaining at least 10 validated segments (the threshold mentioned in Appendix B) over the first 10k iterations would de
  3. Table 2: LingBot-Vision reports ADE20K mIoU of 53.5, trailing the 7B DINOv3 (55.9) by 2.4 points and the distilled DINOv3 ViT-H+ (54.8) by 1.3 points. The abstract claims the model 'matches or surpasses visual foundation models up to seven times larger,' but on semantic segmentation—a core dense spatial task—the gap to DINOv3 is non-trivial. The paper should clarify whether 'matches' refers specifically to depth and video tasks, and temper the headline claim accordingly, or provide additional analysis of why the boundary-centric objective underperforms on semantic segmentation relative to depth.
minor comments (7)
  1. Sec. 3.1, Eq. (4): The iBOT loss uses a single temperature τ without subscript, while Eq. (2) distinguishes τ_t and τ_s. Clarify whether τ in Eq. (4) is the student or teacher temperature.
  2. Sec. 3.3, Eq. (8): The label temperature τ_ℓ is introduced but its value is not specified until Sec. 4.3 ('narrow soft-label temperature'). Report the actual value used.
  3. Table 5, ViT-S row: LingBot-Vision reports KITTI RMSE of 3.784, substantially worse than DINOv2 (3.568) and DINOv3 (2.851). This is mentioned only obliquely ('falls behind on KITTI and Cityscapes'). A brief analysis of why the smallest model degrades specifically on outdoor driving scenes would be helpful.
  4. Fig. 6: The PCA visualizations are compelling but qualitative. Consider adding a quantitative boundary-fidelity metric (e.g., F-measure against BSDS500 or NYUv2 boundary maps) to complement the visual comparison.
  5. Sec. 4.3: The paper states K was reduced from 128 to 32 and 'verified to be quality-neutral.' Report the verification result (e.g., NYUv2 RMSE or ADE20K mIoU at K=128 vs K=32).
  6. Sec. 6.3: The LingBot evaluation set (1,751 frames, 35 scenes) is self-captured with pseudo-ground-truth from FoundationStereo. Clarify whether this set will be publicly released to enable independent verification.
  7. Typos: 'aboundary-centriclens' (abstract), 'Re fined Point Cloud' (Fig. 9 caption), 'DA VIS' should be 'DAVIS' throughout (Tables 3, Sec. 5.2.1).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The three major comments raise legitimate concerns about (1) potential data-distribution leakage from using downstream benchmarks as retrieval seeds, (2) the absence of quantitative evidence for early-training stability of the bootstrapping loop, and (3) the precision of the headline claim given the ADE20K gap to DINOv3. We address each below and commit to revisions in all three cases.

read point-by-point responses

  1. Referee: Sec. 4.1 lists ADE20K, NYU-Depth v2, SUN-RGBD, Cityscapes, and VOC among retrieval seeds; these are the same downstream benchmarks in Tables 2-8. The claim that gains come from the pretraining objective rather than data advantage is not fully controlled. The ImageNet-1K proof of concept (Table 1) controls for data but shows modest gains relative to scaled margins.

    Authors: The referee raises a valid concern. We acknowledge that using downstream evaluation datasets as retrieval seeds creates a distributional alignment that could contribute to the scaled results, and the current manuscript overstates the degree to which the objective alone accounts for the scaled margins. We will revise the manuscript accordingly. revision: yes

  2. Referee: Sec. 3.2, Finding 1: No quantitative metric is provided for how reliably corner anchors produce coherent segments across diverse images, nor for how this reliability evolves during early training. A quantitative ablation over the first 10k iterations would address concerns about bootstrapping stability.

    Authors: The referee is correct that the current evidence for Finding 1 is qualitative only (Fig. 3, Appendix A) and that quantitative metrics would strengthen the claim. We will add the requested quantitative ablation to the revision. revision: yes

  3. Referee: Table 2: LingBot-Vision trails DINOv3 by 2.4 mIoU on ADE20K. The abstract claims 'matches or surpasses' models up to 7x larger, but on semantic segmentation the gap is non-trivial. The paper should clarify or temper the headline claim.

    Authors: The referee is right that the headline claim is imprecise as written. On semantic segmentation, LingBot-Vision does not match the 7B DINOv3; the claim is more accurately scoped to depth and video tasks. We will revise the abstract and introduction to clarify this. revision: yes

Circularity Check

0 steps flagged

No significant circularity: the boundary-centric pretraining derivation is self-contained, with standard self-distillation bootstrapping and external benchmarks

full rationale

The paper's core derivation chain is not circular. The boundary-forcing masked modeling objective (Sec. 3.2, Eq. 6) is defined structurally: boundary tokens B are identified by the teacher's predicted boundary field, forced into the mask, and supervised by a categorical boundary target (Eq. 9). This is a standard self-distillation loop with EMA teacher, not a self-definitional circle. The teacher generates targets online (Sec. 3.4), validated by a parameter-free a-contrario test (Appendix B) whose null hypothesis (uniform orientation distribution) is independently grounded in classical detection theory [14, 47], not defined in terms of the paper's own results. Finding 1 (Sec. 3.2) — that corner points plus level-line orientation produce coherent segments even from random fields — is demonstrated empirically (Fig. 3, Appendix A) as a property of the boundary-field representation, not assumed. The categorical reparameterization (Sec. 3.3, Eq. 8) is a design choice justified by stability arguments, not a fitted parameter renamed as prediction. The ImageNet-1K proof-of-concept (Table 1) controls for data and architecture, isolating the objective's contribution. The scaled results (Table 2) are evaluated against external benchmarks (NYUv2, KITTI, ADE20K, Cityscapes, VOC) with frozen-feature linear probing, an externally falsifiable protocol. The skeptic's concern about downstream datasets appearing as retrieval seeds in the pretraining corpus (Sec. 4.1) is a legitimate data-contamination concern for correctness risk, but it is not circularity: the paper does not define its evaluation metrics in terms of its training data, nor does it fit a parameter to the test set and call it a prediction. The claim that 'gains come from the pretraining objective rather than from a data advantage' (Sec. 4.1) is an inference the paper acknowledges is suggestive ('suggests that...'), not a theorem. Self-citations to prior work on attraction-field representations [57-59] are foundational but not load-bearing for the present paper's claims: the paper re-derives the field properties it needs (Sec. 3.3, Fig. 4) and validates them empirically (Finding 1, Fig. 3). No step in the derivation reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 3 invented entities

The axiom ledger reveals a method with a moderate number of hyperparameters, mostly inherited from standard SSL recipes. The key ad-hoc axiom is Finding 1, which underpins the bootstrapping mechanism. The invented entities are well-motivated and validated within the paper's framework, though the frozen corner detector lacks independent evaluation.

free parameters (8)
  • λ_i (iBOT loss weight) = 1.0
    Set to 1.0 in the final recipe (Sec. 4.3).
  • λ_b (boundary loss weight) = 1.0
    Set to 1.0 in the final recipe (Sec. 4.3).
  • λ_k (KoLeo loss weight) = 0.1
    Set to 0.1 in the final recipe (Sec. 4.3).
  • K (number of bins per field channel) = 32
    Reduced from 128 to 32 for scaling, verified quality-neutral (Sec. 4.3).
  • τ_ℓ (label temperature)
    Mentioned as 'narrow' but exact value not specified in main text.
  • τ_d (support threshold) = 5px
    Set to 5px at field resolution (Appendix A).
  • ε (NFA acceptance threshold) = 1
    Canonical choice from a-contrario theory (Appendix B).
  • Alignment tolerance = π/16
    Stricter than LSD's π/8 (Appendix B).
axioms (4)
  • domain assumption Boundaries and shape discontinuities offer essential cues for perceiving geometric properties.
    Foundational premise of the paper (Abstract, Sec. 1).
  • domain assumption The EMA teacher in self-distillation is always slightly better than the student, providing a usable learning signal.
    Standard SSL assumption, invoked to justify bootstrapping (Sec. 3.4).
  • ad hoc to paper Finding 1: Given corner points, a boundary field with random values decodes into corner-anchored fragments, and coherent segments emerge with level-line guidance.
    The core bootstrapping assumption, demonstrated qualitatively in Fig. 3 and Appendix A but not formally proven.
  • domain assumption The a-contrario null hypothesis (uniformly distributed orientations) holds for non-boundary regions.
    Borrowed from classical detection theory (Appendix B).
invented entities (3)
  • Boundary-bearing tokens independent evidence
    purpose: Tokens identified as containing boundary structure, forced into the student's masked set.
    Validated by the a-contrario test and visualized in Fig. 1.
  • Categorical boundary field independent evidence
    purpose: A per-pixel distribution over discretized bins for distance and orientation, replacing continuous regression.
    Learned and validated online; its stability is demonstrated by successful training at scale.
  • Frozen single-block ViT corner-point detector no independent evidence
    purpose: Seeds corner points for boundary decoding.
    A fixed external component; the paper does not provide evidence of its standalone accuracy or robustness, only that it works within the pipeline.

pith-pipeline@v1.1.0-glm · 34105 in / 2647 out tokens · 557160 ms · 2026-07-07T21:48:16.981135+00:00 · methodology

0 comments
read the original abstract

Dense spatial perception is essential for physical intelligence, where visual systems are expected to recover structured, metric, and actionable representations from pixel observations. Modern visual foundation models tend to prioritize semantic invariance, often at the expense of detailed spatial understanding. In this work, we study vision pretraining through a boundary-centric lens, motivated by the premise that boundaries and shape discontinuities offer essential cues for perceiving geometric properties. Concretely, we propose masked boundary modeling, a self-supervised paradigm that dynamically learns sub-pixel boundary representations and subsequently leverages the discovered boundary-bearing tokens as masked targets to facilitate dense visual token learning. By scaling this framework, we develop LingBot-Vision and demonstrate its efficacy across a diverse set of downstream vision tasks with DINOv3 as a strong baseline. Remarkably, LingBot-Vision drives the progression from LingBot-Depth 1.0 to LingBot-Depth 2.0 for depth completion, and thereby yields enhanced depth estimation, a key pillar for embodied artificial intelligence. Our findings reveal that boundary modeling goes beyond simple line segments and instead serves as a scalable pretraining principle for learning spatially structured visual representations.

Figures

Figures reproduced from arXiv: 2607.05247 by Bin Tan, Changjiang Sun, Kecheng Zheng, Nan Xue, Shaohui Liu, Xing Zhu, Yinghao Xu, Yujun Shen, Zelin Fu.

Figure 1
Figure 1. Figure 1: LingBot-Vision learns dense representations via boundary-centric masked modeling. Each row, from left to right: the input image; the PCA projection of the frozen teacher’s patch tokens; boundary tokens (pink) obtained by a-contrario validation of dense line proposals decoded from the model’s own boundary-field prediction, overlaid on the accumulated response of the validated proposals; and cosine-similarit… view at source ↗
Figure 2
Figure 2. Figure 2: Boundary-centric masked modeling on a toy scene. (a) An input image and its patch grid. (b) Random masking is content-agnostic: most hidden patches are flat and recoverable from context, while the boundary-bearing patches largely escape the mask. (c) Boundary-forcing masking adds every boundary-bearing patch to the random mask, so the structure of the scene is exactly what the student must reconstruct. (d)… view at source ↗
Figure 3
Figure 3. Figure 3: Boundaries emerge from corner points (Finding 1). For two images, we draw five random boundary fields per image, fix the corner points, and decode line segments from each draw (columns). In the bottom two rows, every field channel is sampled uniformly at random: the decoded segments still anchor on the corner points, but stay short and fragmented. In the top two rows, only the direction channel is guided b… view at source ↗
Figure 4
Figure 4. Figure 4: The boundary field at a glance. (a) Image boundaries are modeled as straight line segments; curved boundaries are chains of short segments. (b) The boundary field lifts the sparse segments into a dense map: every pixel near a segment stores its distance d to the segment and three angles (θ, ϕ1 , ϕ2 ) that locate the segment from that pixel; parallel segments share the same orientation color in θ. (c) The e… view at source ↗
Figure 5
Figure 5. Figure 5: Online generation of boundary targets (toy example). Even when the teacher’s boundary field is noisy, as it is early in training (b), the holistic parameterization lets the many weak per-pixel votes in a segment’s support region aggregate into coherent candidate segments anchored at corner points (c), together with spurious candidates such as face diagonals and background chords. The a-contrario test then … view at source ↗
Figure 6
Figure 6. Figure 6: PCA of frozen patch features. The top three PCA components of the patch features are mapped to RGB, computed per image for each model (one model per column). LingBot-Vision (rightmost) resolves objects as coherent regions with crisp boundaries: individual cars and lane structure in the traffic scene, hen silhouettes against the wire fence, the winding contour of the snake, and fine structures such as flowe… view at source ↗
Figure 7
Figure 7. Figure 7: Boundary-token tracking on three videos. Three boundary-token queries are seeded in the first frame of each video (crosses in the left panels) and tracked by cosine similarity of the frozen patch features over each frame’s boundary tokens; the three heatmap panels show the similarity responses of the three queries. Top pair: a manipulation episode seen from a robot’s head camera, where the queries hold the… view at source ↗
Figure 8
Figure 8. Figure 8: Scaling the curated training data. Depth completion quality as the MDM training set grows from 3M to 20M and 150M samples, for the LingBot-Vision and DINOv2 encoder initializations under the same pipeline. Left: mean absolute relative error (REL, lower is better); right: strict threshold accuracy D102 (higher is better). Both initializations improve monotonically with data, but they are not interchangeable… view at source ↗
Figure 9
Figure 9. Figure 9: LingBot-Depth 2.0 on mirror and glass scenes. Each group shows the input RGB, the raw sensor depth and the refined depth for consecutive frames of a captured sequence, together with front and top views of the refined point cloud. The raw depth is missing exactly on the hardest surfaces: window panes, a glass balustrade and reflective floors return no measurements. The completed regions form flat, contiguou… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of depth completion methods. Each scene shows the completed depth map (top) and the corresponding point cloud (bottom) for OMNI-DC, CDMs, LingBot-Depth 1.0 and LingBot-Depth 2.0, next to the RGB frame with ground-truth depth and the raw sensor input. The baselines leave missing regions open, bend planar surfaces, or scatter floating points near depth discontinuities; LingBot-Depth 2… view at source ↗
Figure 11
Figure 11. Figure 11: Corner anchors and a-contrario validation are complementary safeguards. The same boundary field, decoded from the same LingBot-Vision teacher under identical settings, with corner points taken either from the frozen single-block detector (a, b; 920 candidate segments) or derived from the field itself via jloc = 1 − d (c, d; 2,382 candidates). With reliable corner anchors, the vote-aggregation decoding is … 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

67 extracted references · 67 canonical work pages · 11 internal anchors

  1. [1]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  2. [2]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, Xiao...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

    Randall Balestriero and Yann LeCun. LeJEPA: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    BEiT: BERT pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEiT: BERT pre-training of image transformers. InInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  5. [5]

    Revisiting feature prediction for learning visual representations from video.Transactions on Machine Learning Research, 2024, 2024

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video.Transactions on Machine Learning Research, 2024, 2024. 27

    work page 2024

  6. [6]

    Perception encoder: The best visual embeddings are not at the output of the network

    Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi, Jathushan Rajasegaran, Hanoona Rasheed, Junke Wang, Marco Monteiro, Hu Xu, Shiyu Dong, Nikhila Ravi, Daniel Li, Piotr Dollár, and Christoph Feichtenhofer. Perception encoder: The best visual embeddings are not at the output of the network. InAdvances in N...

    work page 2025

  7. [7]

    A computational approach to edge detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 8(6):679–698, 1986

    John Canny. A computational approach to edge detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 8(6):679–698, 1986

    work page 1986

  8. [8]

    TIPSv2: Advancing vision-language pretraining with enhanced patch-text alignment

    Bingyi Cao, Koert Chen, Kevis-Kokitsi Maninis, Kaifeng Chen, Arjun Karpur, Ye Xia, Sahil Dua, Tanmaya Dabral, Guangxing Han, Bohyung Han, Joshua Ainslie, Alex Bewley, Mithun Jacob, René Wagner, Washington Ramos, Krzysztof Choromanski, Mojtaba Seyedhosseini, Howard Zhou, and André Araujo. TIPSv2: Advancing vision-language pretraining with enhanced patch-te...

    work page 2026

  9. [9]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021

    work page 2021

  10. [10]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational Conference on Machine Learning (ICML), 2020

    work page 2020

  11. [11]

    The Cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213–3223, 2016

    work page 2016

  12. [12]

    JAFAR: Jack up any feature at any resolution

    Paul Couairon, Loick Chambon, Louis Serrano, Jean-Emmanuel Haugeard, Matthieu Cord, and Nicolas Thome. JAFAR: Jack up any feature at any resolution. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  13. [13]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009

    work page 2009

  14. [14]

    Meaningful alignments.International Journal of Computer Vision, 40(1):7–23, 2000

    Agnès Desolneux, Lionel Moisan, and Jean-Michel Morel. Meaningful alignments.International Journal of Computer Vision, 40(1):7–23, 2000

    work page 2000

  15. [15]

    Lawrence Zitnick

    Piotr Dollár and C. Lawrence Zitnick. Fast edge detection using structured forests.IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(8):1558–1570, 2015

    work page 2015

  16. [16]

    Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The Pascal visual object classes (VOC) challenge.International Journal of Computer Vision, 88(2):303–338, 2010

    work page 2010

  17. [17]

    Scaling language-free visual representation learning

    David Fan, Shengbang Tong, Jiachen Zhu, Koustuv Sinha, Zhuang Liu, Xinlei Chen, Michael Rabbat, Nicolas Ballas, Yann LeCun, Amir Bar, and Saining Xie. Scaling language-free visual representation learning. InIEEE/CVF International Conference on Computer Vision (ICCV), 2025

    work page 2025

  18. [18]

    Are we ready for autonomous driving? The KITTI vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354–3361, 2012

    work page 2012

  19. [19]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  20. [20]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    work page 2020

  21. [21]

    Learning to parse wireframes in images of man-made environments

    Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 626–635, 2018

    work page 2018

  22. [22]

    Is my Depth Ground-Truth Good Enough? HAMMER -- Highly Accurate Multi-Modal Dataset for DEnse 3D Scene Regression

    HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, Ales Leonardis, and Benjamin Busam. Is my depth ground-truth good enough? HAMMER – highly accurate multi-modal dataset for DEnse 3D scene regression.arXiv preprint arXiv:2205.04565, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    ScaleLSD: Scalable deep line segment detection streamlined

    Zeran Ke, Bin Tan, Xianwei Zheng, Yujun Shen, Tianfu Wu, and Nan Xue. ScaleLSD: Scalable deep line segment detection streamlined. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6327–6336, 2025

    work page 2025

  24. [24]

    Comparison of monocular depth estimation methods using geometrically relevant metrics on the IBims-1 dataset.Computer Vision and Image Understanding, 191:102877, 2020

    Tobias Koch, Lukas Liebel, Marco Körner, and Friedrich Fraundorfer. Comparison of monocular depth estimation methods using geometrically relevant metrics on the IBims-1 dataset.Computer Vision and Image Understanding, 191:102877, 2020

    work page 2020

  25. [25]

    Manipulation as in Simulation: Enabling Accurate Geometry Perception in Robots

    Minghuan Liu, Zhengbang Zhu, Xiaoshen Han, Peng Hu, Haotong Lin, Xinyao Li, Jingxiao Chen, Jiafeng Xu, Yichu Yang, Yunfeng Lin, Xinghang Li, Yong Yu, Weinan Zhang, Tao Kong, and Bingyi Kang. Manipulation as in simulation: Enabling accurate geometry perception in robots.arXiv preprint arXiv:2509.02530, 2025. 28

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  27. [27]

    V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning

    Lorenzo Mur-Labadia, Matthew Muckley, Amir Bar, Mido Assran, Koustuv Sinha, Mike Rabbat, Yann LeCun, Nicolas Ballas, and Adrien Bardes. V-JEPA 2.1: Unlocking dense features in video self-supervised learning.arXiv preprint arXiv:2603.14482, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick L...

    work page 2024

  29. [29]

    Oswald, and Marc Pollefeys

    Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. DeepLSD: Line segment detection and refinement with deep image gradients. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17327–17336, 2023

    work page 2023

  30. [30]

    The 2017 DAVIS Challenge on Video Object Segmentation

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 DA VIS challenge on video object segmentation.arXiv preprint arXiv:1704.00675, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026

    work page 2026

  32. [32]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning (ICML), 2021

    work page 2021

  33. [33]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021

    work page 2021

  34. [34]

    AM-RADIO: Agglomerative vision foundation model – reduce all domains into one

    Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. AM-RADIO: Agglomerative vision foundation model – reduce all domains into one. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  35. [35]

    ImageNet-21K pretraining for the masses

    Tal Ridnik, Emanuel Ben Baruch, Asaf Noy, and Lihi Zelnik-Manor. ImageNet-21K pretraining for the masses. InAdvances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2021

    work page 2021

  36. [36]

    Sajjan, Matthew Moore, Mike Pan, Ganesh Nagaraja, Johnny Lee, Andy Zeng, and Shuran Song

    Shreeyak S. Sajjan, Matthew Moore, Mike Pan, Ganesh Nagaraja, Johnny Lee, Andy Zeng, and Shuran Song. ClearGrasp: 3D shape estimation of transparent objects for manipulation. InIEEE International Conference on Robotics and Automation (ICRA), pages 3634–3642, 2020

    work page 2020

  37. [37]

    Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger

    Thomas Schöps, Johannes L. Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2538–2547, 2017

    work page 2017

  38. [38]

    Indoor segmentation and support inference from RGBD images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. InEuropean Conference on Computer Vision (ECCV), pages 746–760, 2012

    work page 2012

  39. [39]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    work page 2026

  40. [40]

    A 3×3 isotropic gradient operator for image processing

    Irwin Sobel and Gary Feldman. A 3×3 isotropic gradient operator for image processing. Stanford Artificial Intelligence Project, 1968

    work page 1968

  41. [41]

    Lichtenberg, and Jianxiong Xiao

    Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 567–576, 2015

    work page 2015

  42. [42]

    Masked depth modeling for spatial perception

    Bin Tan, Changjiang Sun, Xiage Qin, Hanat Adai, Zelin Fu, Tianxiang Zhou, Han Zhang, Yinghao Xu, Xing Zhu, Yujun Shen, and Nan Xue. Masked depth modeling for spatial perception. InEuropean Conference on Computer Vision (ECCV), 2026

    work page 2026

  43. [43]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features.arX...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    DIODE: A Dense Indoor and Outdoor DEpth Dataset

    Igor Vasiljevic, Nicholas I. Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A dense indoor and outdoor DEpth dataset.arXiv preprint arXiv:1908.00463, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1908

  45. [45]

    Shashanka Venkataramanan, Valentinos Pariza, Mohammadreza Salehi, Lukas Knobel, Spyros Gidaris, Elias Ramzi, Andrei Bursuc, and Yuki M. Asano. Franca: Nested Matryoshka clustering for scalable visual representation learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    work page 2026

  46. [46]

    Marigold-DC: Zero-shot monocular depth completion with guided diffusion

    Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-DC: Zero-shot monocular depth completion with guided diffusion. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 5359–5370, 2025

    work page 2025

  47. [47]

    LSD: A fast line segment detector with a false detection control.IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(4):722–732, 2010

    Rafael Grompone von Gioi, Jérémie Jakubowicz, Jean-Michel Morel, and Gregory Randall. LSD: A fast line segment detector with a false detection control.IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(4):722–732, 2010

    work page 2010

  48. [48]

    MoGe: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision

    Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. MoGe: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5261–5271, 2025

    work page 2025

  49. [49]

    InternVideo2: Scaling foundation models for multimodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Jilan Xu, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. InternVideo2: Scaling foundation models for multimodal video understanding. InEuropean Conference on Computer Vision (ECCV), 2024

    work page 2024

  50. [50]

    Depth Anything with Any Prior

    Zehan Wang, Siyu Chen, Lihe Yang, Jialei Wang, Ziang Zhang, Hengshuang Zhao, and Zhou Zhao. Depth Anything with any prior.arXiv preprint arXiv:2505.10565, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [51]

    Mapillary street-level sequences: A dataset for lifelong place recognition

    Frederik Warburg, Søren Hauberg, Manuel López-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2623–2632, 2020

    work page 2020

  52. [52]

    FoundationStereo: Zero-shot stereo matching

    Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. FoundationStereo: Zero-shot stereo matching. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5249–5260, 2025

    work page 2025

  53. [53]

    Google Landmarks Dataset v2 - a large-scale benchmark for instance-level recognition and retrieval

    Tobias Weyand, André Araújo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2 - a large-scale benchmark for instance-level recognition and retrieval. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2572–2581, 2020

    work page 2020

  54. [54]

    Unsupervised depth completion from visual inertial odometry

    Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2):1899–1906, 2020

    work page 1906

  55. [55]

    Holistically-nested edge detection

    Saining Xie and Zhuowen Tu. Holistically-nested edge detection. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 1395–1403, 2015

    work page 2015

  56. [56]

    Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas S. Huang. YouTube-VOS: A large-scale video object segmentation benchmark.arXiv preprint arXiv:1809.03327, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  57. [57]

    Learning attraction field representation for robust line segment detection

    Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, and Liangpei Zhang. Learning attraction field representation for robust line segment detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1595–1603, 2019

    work page 2019

  58. [58]

    Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip H. S. Torr. Holistically-attracted wireframe parsing. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2785–2794, 2020

    work page 2020

  59. [59]

    Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip H. S. Torr. Holistically-attracted wireframe parsing: From supervised to self-supervised learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14727–14744, 2023

    work page 2023

  60. [60]

    In pursuit of pixel supervision for visual pre-training.arXiv preprint arXiv:2512.15715, 2025

    Lihe Yang, Shang-Wen Li, Yang Li, Xinjie Lei, Dong Wang, Abdelrahman Mohamed, Hengshuang Zhao, and Hu Xu. In pursuit of pixel supervision for visual pre-training.arXiv preprint arXiv:2512.15715, 2025

    work page arXiv 2025

  61. [61]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    work page 2023

  62. [62]

    Scene parsing through ADE20K dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5122–5130, 2017

    work page 2017

  63. [63]

    DINO-WM: World models on pre-trained visual features enable zero-shot planning

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. DINO-WM: World models on pre-trained visual features enable zero-shot planning. InInternational Conference on Machine Learning (ICML), 2025. 30

    work page 2025

  64. [64]

    iBOT: Image BERT pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pre-training with online tokenizer. InInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  65. [65]

    End-to-end wireframe parsing

    Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 962–971, 2019

    work page 2019

  66. [66]

    Any to full: Prompting Depth Anything for depth completion in one stage.arXiv preprint arXiv:2603.05711, 2026

    Zhiyuan Zhou, Ruofeng Liu, Taichi Liu, Weijian Zuo, Shanshan Wang, Zhiqing Hong, and Desheng Zhang. Any to full: Prompting Depth Anything for depth completion in one stage.arXiv preprint arXiv:2603.05711, 2026

    work page arXiv 2026

  67. [67]

    OMNI-DC: Highly robust depth completion with multiresolution depth integration

    Yiming Zuo, Willow Yang, Zeyu Ma, and Jia Deng. OMNI-DC: Highly robust depth completion with multiresolution depth integration. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 9287–9297, 2025. 31

    work page 2025