Pith.

Computer Vision and Pattern Recognition · arXiv:2610.26625 v1 · submitted 19 Oct 2026 · 78 references

EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience

Xinhao Ji, Hui En Pang, Hao Li, Kairui Hu, Jiner Wang, Mingze Sun, Ziang Cao, Yalun Dai

show 7 more authors

Thang Luu, Qin Ma, Hongyi Chen, Yukang Cao, Fangzhou Hong, Zhaoxi Chen, Ziwei Liu

The paper says

“Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer's motion, surrounding geometry, and hand-object interactions must fit together in a consistent world.”

more of the abstract

Achieving this coherence requires both joint reconstruction and aligned supervision, yet existing methods and datasets offer limited support for either. We present EgoXperience, a full-stack paradigm that addresses these needs through data collection, automatic annotation, unified modeling, and benchmarking. Central to EgoXperience is an automatic annotation engine, EgoX-Engine, that couples geometrically verified object tracking with SLAM-based world alignment, yielding consistent supervision for the camera, scene, hands, and objects. This yields EgoX-Data-6M, a 50-hour dataset of task-driven manipulation and locomotion within and across indoor and outdoor environments, with second-level task captions. We carefully select one hour across these activity categories as EgoX-Bench to evaluate joint reconstruction. To translate this supervision into reconstruction, we introduce EgoX, trained with EgoX-Data-6M and complementary annotations from additional public datasets. Given monocular egocentric video and first-frame object masks, our model achieves joint metric reconstruction of the camera, scene, hands, and objects in a shared world coordinate frame with a single forward pass. Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery. Against the strongest evaluated baselines, EgoX reduces camera, hand, and depth errors by 68.9%, 65.6%, and 42.9%, and improves object pose scores by 28.0%.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 20 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Substantial dataset-plus-model systems paper whose external-benchmark results hold up; the self-annotated EgoX-Bench gains are real but should be read as upper bounds until the engine gets an independent audit. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 78 references say.

The paper says, at Sec. 3.2, Tables 3 and 5

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

EgoX-Bench is called 'quality-controlled' (Sec. 3.2), but the criterion is never operationalized: the one-hour subset is annotated by the same EgoX-Engine that generated the EgoX-Data-6M training supervision, and the text gives no thresholds for the FoundationPose depth-verification pass, no hand-fit or SLAM drift audit, no selection protocol, and no per-clip error statistics.

more of objection 1

The two largest gains in the abstract, 65.6% hand W-MPJPE and 42.9% depth AbsRel, are computed on EgoX-Bench, so a systematic engine bias (SLAM scale drift, MANO fit drift, or depth-verification failure during occlusion) shared by the labels and by a model trained on those labels would inflate exactly the headline arithmetic. The external benchmarks keep the broader claim plausible, but the EgoX-Bench numbers themselves need an independent check. I ask for at least one of the following: a human or multi-camera audit on a sampled subset of EgoX-Bench with quantitative disagreement statistics; reported pass-rates and residual distributions for the engine's verification step; or a revision that demotes the EgoX-Bench-only percentages in the abstract in favor of the external-benchmark results. Relatedly, the paper should state whether the 101 object instances in EgoX-Data-6M overlap with those in EgoX-Bench: sequence-level exclusion (Sec. 4) does not prevent instance-level memorization, which could make the object and hand benchmarks easier for EgoX than for baselines that never saw the objects.

The paper says, at Sec. 4.1, Appendix C, Table 4

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

Object 6DoF poses are evaluated only in camera coordinates (Appendix C), so the central claim that EgoX reconstructs camera, scene, hands, and objects at metric scale in a shared world coordinate frame is never tested for the object component. A model with accurate camera-frame object poses but inconsistent composition with its own predicted trajectory would pass the current object metrics unchanged.

more of objection 2

I recommend reporting world-frame ADD/ADD-S using each method's estimated camera trajectory, the direct analogue of the W-MPJPE metric already used for hands; this would exercise the composition that the paper's headline claim rests on.

5 minor objections

The paper says, at Abstract and Sec. 1

Pith says: minor objection 1 of 5

The four headline reductions (68.9%, 65.6%, 42.9%, 28.0%) are reported without the benchmark each refers to (HOI4D, EgoX-Bench, EgoX-Bench, and HOT3D, respectively); stating the benchmark alongside each percentage would prevent misreading.

The paper says, at Author block and Fig. 1

Pith says: minor objection 2 of 5

The author/affiliation block contains garbled text ('Y ukang Cao', 'D Qin Ma 1,D', 'Thang Luu 1,D'), and the inline numeric annotations inside Fig. 1 are difficult to read at figure scale; both should be cleaned.

show 3 more

The paper says, at Table 5

Pith says: minor objection 3 of 5

The dagger symbol on the 'Depth' column header is never footnoted; either add the footnote or remove the symbol.

The paper says, at Sec. 3.3 and Tables 6-7

Pith says: minor objection 4 of 5

The coupled geometric supervision L_coupled is the mechanism claimed to tie camera, hand, and object estimation together, but the ablations test only ScaleHead and DepthHead, and only on HOT3D. An ablation removing L_coupled, ideally also reported on an external benchmark, would directly support the jointness claim that distinguishes EgoX from modular baselines.

The paper says, at Appendix C

Pith says: minor objection 5 of 5

Please state whether the matched 150-observation clips are contiguous windows or sampled frames, since this affects the interpretation of the nATE and W-MPJPE numbers across methods.

What would settle it

Select a random 10 to 15 minute subset of EgoX-Bench clips not used for model development, re-annotate them with an independent protocol (for example, manually verified multi-camera 6DoF in a calibrated rig or a second marker-based tracking system), and recompute the EgoX-Bench rows of Tables 3 to 5 using those labels. If W-MPJPE, AbsRel, or AUC-S50 move by more than about 10 percent relative, the EgoX-Bench numbers overstate accuracy; if they are stable, the engine-bias concern does not land. The subset selection should be pre-registered to avoid re-curating the benchmark.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 20 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 1 min 20 s · 22,225 tokens read, 10,973 written
  3. Editor wrote the verdict above 30 s · 23,259 tokens read, 4,102 written
  4. Referee major revision: 2 major objections, 5 minor 57 s · 22,518 tokens read, 8,646 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The short version: this is a real contribution, and the central claim survives contact. The 50-hour dataset is, as far as Table 1 shows, the first egocentric resource to combine cross-room and outdoor capture with aligned depth, hand, object, and mask annotations in a common metric frame. The model is the first end-to-end feed-forward method to output cameras, depth, hands, and objects together, and the authors ship code, data, and models, so the resource is checkable.

What I verified: the reduction arithmetic in the abstract checks out (328.81 to 113.24 mm is 65.6% W-MPJPE; 0.0806 to 0.0460 is 42.9% AbsRel). The paper is also honest where it would be easy to hype: VGGT-Ω beats EgoX on camera on HOT3D and EgoX-Bench, and the text says so. The external benchmarks carry the argument. On HOT3D, H2O, and HOI4D — ground truth that did not go through their engine — EgoX wins hand and object metrics over strong baselines that actually receive more inputs (metric CAD assets, per-frame masks, fitted depth). That is evidence the joint model is genuinely better, not merely tuned to its own labels.

The soft spot is the one the stress test named, and it is real: EgoX-Bench is annotated by the same EgoX-Engine that produced the training supervision, and the paper never defines “quality-controlled,” gives no selection protocol for the one-hour subset, and reports no error bars. The two headline-cited gains (65.6% hand, 42.9% depth) live on that benchmark, so those specific percentages should not be taken at face value. What mitigates this is that the same design is corroborated, at smaller margins, on independently labeled data. So this is a conditional concern rather than a load-bearing flaw: it inflates the in-house numbers, not the overall claim.

Minor and acknowledged: rigid objects only, first-frame object masks required, and the VGGT/VGGT-Ω depth comparison uses a fitted scale per sequence. They disclose all of these; no penalty.

Who this is for: egocentric 3D, hand-object reconstruction, and anyone building automatic annotation engines for embodied data. It deserves a serious referee. The referee's main job is the annotation engine: ask for a human or multi-camera audit of a sample of EgoX-Bench labels, the exact subset-selection rule, and run-to-run variance. I would send it to review and expect that revision to land.

The strongest objection

The load-bearing premise is that the automatic annotations produced by EgoX-Engine are accurate enough to define both training targets and benchmark ground truth. The engine (Sec. 3.1) obtains object 6DoF from FoundationPose tracking verified by depth agreement, hands from stereo triangulation plus MANO fitting, and camera alignment from SLAM; EgoX-Bench is annotated by the same pipeline (Sec. 3.2). If any component has a systematic bias, for example SLAM scale drift, depth-verification failure during occlusion, or MANO hand-fit drift, a model trained on the same labels can learn to reproduce that bias and is then rewarded on EgoX-Bench because the benchmark uses the identical labels as ground truth. The paper does not define what quality-controlled means for the one-hour subset, does not report an independent human or multi-camera audit of the engine output, and gives no selection protocol or error bars, so the headline EgoX-Bench gains (W-MPJPE 113.24 mm, AbsRel 0.0460, AUC-S50 76.59 in Tables 3 to 5) cannot be separated from engine bias on the basis of the paper alone. This is the most load-bearing concern because the metric-scale and world-grounded claim is precisely where annotation scale and drift matter most: all components are aligned through the same camera trajectory, so a single systematic drift can inflate camera, hand, depth, and object results simultaneously. The external benchmarks (HOT3D, H2O, HOI4D) have independent ground truth and keep the broader claim plausible, so the appropriate outcome is conditional acceptance pending independent verification, not rejection.

The paper's central claim rests on two pillars: the EgoX-Data-6M resource with its automatic annotation engine, and the EgoX model that is trained and evaluated on those annotations plus public datasets. The model's external-benchmark results are strong, and the reader's arithmetic checks hold, so the weakest point is the dependence of both training and evaluation on the same unverified automatic labels. This is a verification risk rather than a disagreement with current consensus: the engine's SLAM, FoundationPose, and MANO components are all reasonable tools, but no independent check is described that would catch a shared systematic error. I considered the metric-scale ambiguity of monocular reconstruction and the single-forward-pass claim as alternative concerns, but they are either learnable in principle or an architectural statement; biased EgoX-Bench labels would directly corrupt the headline numbers in Tables 3 to 5. The authors' stated rigid-object limitation is acknowledged and does not affect this analysis. Because HOT3D, H2O, and HOI4D provide independent ground truth and support the broader model claim, the concern warrants conditional acceptance with an independent audit rather than a rejection or a stronger accusation. The reader's conditional verdict therefore remains the right outcome, and the concrete test above would settle whether the EgoX-Bench-specific concern actually lands.

The referee's report

If the results hold, this is a genuine advance: a single feed-forward model replaces the usual composition of SLAM, monocular depth, hand, and object-pose systems, with fewer auxiliary inputs (no ground-truth cameras, scale, or per-frame masks in the main comparisons) while retaining metric scale and world-frame consistency. I credit several concrete strengths. First, the headline arithmetic is reproducible: I checked the 68.9%, 65.6%, 42.9%, and 28.0% reductions against Tables 3-5 and they are correct. Second, the evaluation protocol is unusually transparent: scale-fitting for VGGT/VGGT-Omega is disclosed, baseline auxiliary inputs are enumerated per method, splits are sequence-disjoint, and the paper states explicitly that no method receives ground-truth cameras in the main comparisons. Third, code, data, and models are released. Fourth, the external benchmarks (HOT3D, H2O, HOI4D) provide independent ground truth, so the broad claim of state-of-the-art joint reconstruction is anchored outside the self-annotated data.

Pith's reading of the paper (a machine's summary, not the authors')

EgoXperience claims that coherent 3D understanding of egocentric activity — the wearer's camera motion, the surrounding scene, and the hands and objects in it — should be reconstructed all at once in a shared metric world frame, and that this is now achievable in a single feed-forward pass rather than by stitching together SLAM, depth, hand, and object pose systems. The paper's central assertion is that the real bottleneck was supervision: every component must be annotated in one consistent coordinate system, which existing datasets rarely provide at scale. It therefore builds EgoX-Engine, an automatic annotation pipeline that couples geometrically verified object tracking with SLAM-based world alignment, producing a 50-hour dataset (EgoX-Data-6M) and a one-hour held-out benchmark (EgoX-Bench). The model, EgoX, is trained with a loss that composes predicted hand and object geometry through the predicted cameras, so the components are co-supervised through their consistency. If the claim holds, embodied agents can obtain a coherent metric world from a wearable camera without integrating several error-prone modules.

What carries the argument. The mechanism has two halves. On the data side, EgoX-Engine couples FoundationPose object tracking, verified by agreement between the tracked object model and observed depth, with SLAM camera trajectories from a calibrated head rig, transforming scene, hand, and object annotations into one session-fixed metric frame so that the supervision itself is geometrically consistent. On the model side, EgoX is a Transformer in which DINOv3 patch features are joined by camera, hand, object, and scale tokens; heads read this shared representation, and the composition rule $\mathrm{U}_t(\tilde{u}, z) = \hat{R}_t^\top (z \hat{K}_t^{-1} \tilde{u} - \hat{\tau}_t)$ maps every image position and predicted depth into the common world frame $\mathcal{W}$, with a single predicted scale factor $\hat{s}$ applied consistently to scene, wrist, and object positions. The coupled loss evaluates the composed hand-object geometry in $\mathcal{W}$ and backpropagates through the camera transform, which is what ties camera estimation to hand and object accuracy.

Core claim. The load-bearing claim is stated in Section 3.3: given monocular egocentric video and first-frame object masks, EgoX jointly reconstructs the camera, scene depth, hands, and objects at metric scale in a shared world coordinate system, all in a single forward pass. All outputs are expressed in the coordinate frame of the first camera, and one shared scale head converts the whole window to metric units. During training, predicted hand and object geometry is transformed through the estimated camera poses into the world frame and supervised there, so errors in camera estimation are penalized through the misplacement of hands and objects. The paper reports that EgoX achieves the best depth, hand, and object metrics on all four benchmarks (HOT3D, EgoX-Bench, HOI4D, H2O) and the best camera accuracy on HOI4D, with headline numbers such as world-space hand error (W-MPJPE) of 113.24 mm versus 328.81 mm for WiLoR plus SLAM on EgoX-Bench, depth error (AbsRel) of 0.0460, and object pose accuracy (AUC-S50) of 76.59 against 59.80 for a baseline that additionally receives metric CAD models and per-frame masks.

The premise it rests on. The load-bearing premise is that EgoX-Engine's automatic labels — object poses checked by depth agreement, hands fitted by triangulation plus parametric MANO fitting, and cameras from SLAM — are accurate enough to serve as ground truth for both training and the EgoX-Bench evaluation; if the engine drifts systematically, the benchmark will reward models that reproduce that same drift.

If the paper is right (Pith's inferences, not the authors')

A single end-to-end model can replace the usual composition of SLAM, monocular depth, hand, and object-tracker pipelines for egocentric reconstruction, using fewer inputs: no CAD assets, no per-frame masks, and no ground-truth scale.

Because the training objective applies each loss only where labels exist, partially annotated public datasets contribute to the same coupling mechanism, so the approach can absorb new data without full annotation coverage.

The ablations indicate the coupling does real work: removing the scale head or the depth head degrades hand and object accuracy even when an oracle scale is supplied, meaning joint supervision changes what each component learns rather than merely combining them.

Automatic world-aligned annotation can scale to new indoor and outdoor settings without motion-capture or calibrated multi-camera infrastructure, as long as stereo depth, SLAM trajectories, and pre-scanned object meshes are available.

The results on HOT3D, H2O, and HOI4D, whose ground truth comes from capture pipelines different from the authors' engine, suggest the learned joint representation transfers across annotation conventions and is not confined to the authors' own benchmark.

A decisive next check, not performed in the paper, is evaluation on ground truth produced by a physically independent method — a scanned scene with camera and object trajectories recovered by a calibrated multi-view or motion-capture rig — since EgoX-Bench is labeled by the same engine that produced the training supervision.

The engine requires pre-scanned metric object meshes built in a calibrated dome, which is the main scalability constraint; a natural extension is reconstructing object geometry in situ from the egocentric video itself and feeding it back through the verified-tracking loop.

The single shared scale factor per window is a strong inductive bias that real long recordings may stress; a testable extension is whether one factor per window remains sufficient on hour-long outdoor sequences where scale drift usually accumulates.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer's motion, surrounding geometry, and hand-object interactions must fit together in a consistent world. Achieving this coherence requires both joint reconstruction and aligned supervision, yet existing methods and datasets offer limited support for either. We present EgoXperience, a full-stack paradigm that addresses these needs through data collection, automatic annotation, unified modeling, and benchmarking. Central to EgoXperience is an automatic annotation engine, EgoX-Engine, that couples geometrically verified object tracking with SLAM-based world alignment, yielding consistent supervision for the camera, scene, hands, and objects. This yields EgoX-Data-6M, a 50-hour dataset of task-driven manipulation and locomotion within and across indoor and outdoor environments, with second-level task captions. We carefully select one hour across these activity categories as EgoX-Bench to evaluate joint reconstruction. To translate this supervision into reconstruction, we introduce EgoX, trained with EgoX-Data-6M and complementary annotations from additional public datasets. Given monocular egocentric video and first-frame object masks, our model achieves joint metric reconstruction of the camera, scene, hands, and objects in a shared world coordinate frame with a single forward pass. Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery. Against the strongest evaluated baselines, EgoX reduces camera, hand, and depth errors by 68.9%, 65.6%, and 42.9%, and improves object pose scores by 28.0%.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-20

The authors' main claim, as the classifier picked it out 0.59

“Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.15 · human participants 0.70 · controlled comparison 0.89 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.11

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer's motion, surrounding geometry, and hand-object interactions must fit together in a consistent world. background 0.97 0.11 0.08 asserted 1.00
Achieving this coherence requires both joint reconstruction and aligned supervision, yet existing methods and datasets offer limited support for either. gap 0.99 0.06 0.04 asserted 1.00
We present EgoXperience, a full-stack paradigm that addresses these needs through data collection, automatic annotation, unified modeling, and benchmarking. contribution 1.00 0.92 0.14 asserted 1.00
Central to EgoXperience is an automatic annotation engine, EgoX-Engine, that couples geometrically verified object tracking with SLAM-based world alignment, yielding consistent supervision for the camera, scene, hands, and objects. method 0.91 0.93 0.08 asserted 1.00
This yields EgoX-Data-6M, a 50-hour dataset of task-driven manipulation and locomotion within and across indoor and outdoor environments, with second-level task captions. contribution 0.64 0.95 0.04 asserted 1.00
We carefully select one hour across these activity categories as EgoX-Bench to evaluate joint reconstruction. method 0.72 0.96 0.03 asserted 0.99
To translate this supervision into reconstruction, we introduce EgoX, trained with EgoX-Data-6M and complementary annotations from additional public datasets. method 0.85 0.94 0.06 asserted 0.99
Given monocular egocentric video and first-frame object masks, our model achieves joint metric reconstruction of the camera, scene, hands, and objects in a shared world coordinate frame with a single forward pass. method 0.77 0.95 0.28 asserted 1.00
Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery. result 0.96 0.95 0.36 asserted 1.00
Against the strongest evaluated baselines, EgoX reduces camera, hand, and depth errors by 68.9%, 65.6%, and 42.9%, and improves object pose scores by 28.0%. result 1.00 0.96 0.47 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.99 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-20: 1 request, 7,743 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (10)
Figure 2 of the paper
Figure 2. Overview of EgoX-Data-6M. Examples and statistics summarize capture settings, task coverage, and aligned multimodal annotations. Capture and geometric reference. We record with the HOMIE Gen2 head-mounted rig [54], using its associated processing pipeline to estimate metric depth from the synchronized, cali￾brated front stereo pair and recover camera trajectories via SLAM from its four calibrated fisheye cameras. For each recording, we use the SLAM coordinate system as a fixed session frame and transform camera-relative scene, hand, and object estimates into this frame. Object…
Figure 3 of the paper
Figure 3. Overview of EgoX-Engine. The pipeline converts egocentric recordings into structured multimodal annotations, including aligned geometry and temporally localized task captions. 3.2. EGOX-Data-6M Given EGOX-Engine, we apply it to annotate task-driven egocentric captures, constructing EGOX-Data-6M, which comprises 50 hours of video, 101 rigid objects, and 270 task scripts. The recordings cover activities ranging from manipulation within a single workspace to interactions across rooms and loco-manipulation in outdoor environments. The engine provides geometri￾cally aligned camera,…
Figure 4 of the paper
Figure 4. Overview of EgoX. Shared scene and entity features support joint reconstruction, while predicted cameras and scale place all outputs in a common metric world frame. RGB HaWoR MINT ReViV WiLoR + SLAM HandFlow EgoX (Ours) Ground Truth H O T 3 D E g o X-B e n c h
Figure 5 of the paper
Figure 5. World-Space Hand and Camera Reconstruction. Qualitative comparisons on HOT3D and EgoX-Bench. Predictions and ground truth are shown in a common world frame. 3.3. EGOX To translate the aligned supervision into joint reconstruction, we propose EGOX, which is capable of predicting the camera, scene, and motions of both hands and objects within a single model. The central challenge to achieve this goal is that the predictions must agree geometrically as both the viewpoint and interaction change. As illustrated in
Figure 6 of the paper
Figure 6. Camera-Frame Object Pose Reconstruction. Qualitative comparisons of object 6-DoF predic￾tions with ground truth on HOT3D and EgoX-Bench. RGB VGGT VGGT-Ω DA3 Nested MapAnything ReViV EgoX (Ours) GT depth H O T 3 D E g o X-B e n c h
Figure 7 of the paper
Figure 7. Depth Reconstruction. GT: refined official depth. VGGT/VGGT-Ω: one scale fitted to DA3 Nested depth per sequence; others: predicted metric scale. Joint scene and entity representation. To jointly estimate cameras, scene geometry, and hand￾object motion, we encode them with a shared Transformer. Starting from image patch features extracted by DINOv3 [55], we introduce camera, hand, and object tokens in each view, using distinct left/right embeddings to identify the hands and first-frame mask descriptors to associate target objects across views. These tokens gather image evidence…
Figure 8 of the paper
Figure 8. Additional World-Space Hand and Camera Reconstruction. Qualitative comparisons on HOI4D and H2O, supplementing
Figure 10 of the paper
Figure 10. Additional Depth Reconstruction Results. Comparisons use the same reference-depth processing and scale protocol as

Showing 8 of the paper's 10 figures. Figures are the authors', reproduced from the paper.

References (78)
  1. [1] AgiBot World Team. 2026. Agibot world 2026. https://huggingface.co/datasets/agibot-world/ AgiBotWorld2026 on Pith
  2. [2] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. 2024. Scenescript: Reconstructing scenes with an autor... on PithPith's reviewarXiv
  3. [3] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Shangchen Han, Fan Zhang, Linguang Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, Jakob Julian Engel, and Tomas Hodan. 2025. Hot3d: Hand and object tracking in 3d from egoce... on Pith
  4. [4] Siddhant Bansal, Zhifan Zhu, Shashank Tripathi, Jiahe Zhao, Michael J. Black, and Dima Damen. 2026. Towards in-the-wild egocentric 3d hand-object pose estimation. InEuropean Conference on Computer Vision (ECCV) on Pith
  5. [5] Yukang Cao, Xinying Guo, Mingyuan Zhang, Haozhe Xie, Chenyang Gu, and Ziwei Liu. 2026. Crowdmogen: Event-driven collective human motion generation.International Journal of Computer Vision, 134(1):29 on Pith
  6. [6] Yukang Cao, Haozhe Xie, Fangzhou Hong, Long Zhuo, Zhaoxi Chen, Liang Pan, and Ziwei Liu. 2026. Hsimul3r: Physics-in-the-loop reconstruction of simulation-ready human–scene interactions. InEuropean Conference on Computer Vision, pages 509–528. Springer on Pith
  7. [7] Yukang Cao, Haozhe Xie, Beichen Wen, Runmao Yao, Yinghao Liu, Yue Huang, Zhichao Liao, Yunxiang Wang, Haiheng Liu, Xingshun Tian, Dawei Su, Long Zhuo, Dacheng Tao, Xiaogang Wang, Liang Pan, and Ziwei Liu on Pith
  8. [8] Ziang Cao, Zhaoxi Chen, Liang Pan, and Ziwei Liu. 2025. Physx-3d: Physical-grounded 3d asset generation.arXiv preprint arXiv:2507.12465 on PitharXiv
  9. [9] Ziang Cao, Fangzhou Hong, Zhaoxi Chen, Liang Pan, and Ziwei Liu. 2025. Physx-anything: Simulation-ready physical 3d assets from single image.arXiv preprint arXiv:2511.13648 on PitharXiv
  10. [10] Ziang Cao, Yinghao Liu, Haitian Li, Runmao Yao, Fangzhou Hong, Zhaoxi Chen, Liang Pan, and Ziwei Liu on Pith
  11. [11] Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman Rädle, and 19 o... on PithPith's reviewarXiv
  12. [12] Physx-omni: Unified simulation-ready physical 3d generation for rigid, deformable, and articulated objects. Preprint, arXiv:2605.21572 on PithPith's reviewarXiv
  13. [13] Yuantao Chen, Jiahao Chang, Chongjie Ye, Chaoran Zhang, Zhaojie Fang, Chenghong Li, and Xiaoguang Han. 2026. Forehoi: Feed-forward 3d object reconstruction from daily hand-object interaction videos.Preprint, arXiv:2602.06226 on PithPith's reviewarXiv
  14. [14] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, Jan Kautz, and Dieter Fox. 2021. DexYCB: A benchmark for capturing hand grasping of objects. InIEEE/CVF Conference on Computer Vision... on Pith
  15. [15] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. 2022. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100.Intern... on Pith
  16. [16] Yue Chen, Xingyu Chen, Yuxuan Xue, Anpei Chen, Yuliang Xiu, and Pons-Moll Gerard. 2025. Human3r: Everyone everywhere all at once.arXiv preprint arXiv:2510.06219 on PitharXiv
  17. [17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. 2023. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. InProceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  18. [18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Muhammed Kocabas, Xu Chen, Michael J Black, and Otmar Hilliges. 2024. HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patter... on Pith
  19. [19] Hongming Fu, Wenjia Wang, Xiaozhen Qiao, Rolandos Alexandros Potamias, Taku Komura, Shuo Yang, Zheng Liu, and Bo Zhao. 2026. Egograsp: World-space hand-object interaction estimation from egocentric videos. Preprint, arXiv:2601.01050 on PitharXiv
  20. [20] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J. Black, Trevor Darrell, and Angjoo Kanazawa. 2025. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. InProceedings of the IEEE/CVF International Conference on Computer Vision on Pith
show 58 more
  1. [21] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zach Chavis, Joya Chen, Feng Cheng, Fu-Jen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, Maria Esc... on Pith
  2. [22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Ku- mar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Meng... on Pith
  3. [23] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. 2022. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. InIEEE Computer Vision and Pattern Recognition Conference on Pith
  4. [24] Ryan Hoque, Peide Huang, David J. Yoon, Mouli sivapurapu, and Jian Zhang. 2026. Egodex: Learning dex- terous manipulation from large-scale egocentric video. InThe Fourteenth International Conference on Learning Representations on Pith
  5. [25] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. 2020. HOnnotate: A method for 3d annotation of hand and object poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition on Pith
  6. [26] Jiahui Huang, Qunjie Zhou, Hesam Rabeti, Aleksandr Korovko, Huan Ling, Xuanchi Ren, Tianchang Shen, Jun Gao, Dmitry Slepichev, Chen-Hsuan Lin, Jiawei Ren, Kevin Xie, Joydeep Biswas, Laura Leal-Taixe, and Sanja Fidler. 2025. Vipe: Video pose engine for 3d geometric perception.P... on PithPith's reviewarXiv
  7. [27] Zhi Jing, Siyuan Yang, Jicong Ao, Ting Xiao, Yu-Gang Jiang, and Chenjia Bai. 2025. Humanoidgen: Data generation for bimanual dexterous manipulation via LLM reasoning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems on Pith
  8. [28] Wendi Hu, Haonan Zhou, Wenhao Hu, and Gaoang Wang. 2026. Hand3r: Online 4d hand-scene reconstruction in the wild.Preprint, arXiv:2602.03200 on PitharXiv
  9. [29] Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, Jonathon Luiten, Manuel Lopez-Antequera, Samuel Rota Bulò, Christian Richardt, Deva Ramanan, Sebastian Scherer, and Peter... on Pith
  10. [30] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. 2021. H2o: Two hands manipulating objects for first person interaction recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10138–10148 on Pith
  11. [31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht on Pith
  12. [32] Taeyeop Lee, Bowen Wen, Minjun Kang, Gyuree Kang, In So Kweon, and Kuk-Jin Yoon. 2025. Any6D: Model-free 6d pose estimation of novel objects. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 13 on Pith
  13. [33] Baicheng Li, Dong Wu, Jun Li, Shunkai Zhou, Zecui Zeng, Lusong Li, and Hongbin Zha. 2026. Mv-sam3d: Adaptive multi-view fusion for layout-aware 3d generation.arXiv preprint arXiv:2603.11633 on PithPith's reviewarXiv
  14. [34] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martín-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, Hang Yin, Michael Lingelbach, Minjune Hwang, Ayano Hiranaka, Sujay Garlanka, Arman Aydin, Sharon Lee, Jiankai Sun, Mona Anvari... on PithPith's reviewarXiv
  15. [35] Yann Labbé, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. 2022. Megapose: 6d pose estimation of novel objects via render & compare. InProceedings of the 6th Conference on Rob... on Pith
  16. [36] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavely. 2025. MegaSaM: Accurate, Fast and Robust Structure and Motion from Casual Dynamic Videos. InProceedings of the IEEE/CVF Conference on Comput... on Pith
  17. [37] Haotong Lin, Sili Chen, Junhao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang on Pith
  18. [38] Yumeng Liu, Xiao-Xiao Long, Marc Habermann, Xuanze Yang, Cheng Lin, Yuan Liu, Yuexin Ma, and Ligang Liu. 2026. Hggt: Robust and flexible 3d hand mesh reconstruction from uncalibrated images.arXiv preprint arXiv:2603.23997 on PithPith's reviewarXiv
  19. [39] Hao Li, Zhengyu Zou, Fangfu Liu, Xuanyang Zhang, Fangzhou Hong, Yukang Cao, Yushi LAN, Manyuan Zhang, Gang YU, Dingwen Zhang, and Ziwei Liu. 2026. IGGT: Instance-grounded geometry transformer for semantic 3d reconstruction. InThe Fourteenth International Conference on Learning... on Pith
  20. [40] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. 2022. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio... on Pith
  21. [41] Xiaozhong Lyu, Gen Li, Zhiyin Qian, Xucong Zhang, Marc Pollefeys, and Siyu Tang. 2026. Reviv: Reconstructing the viewer and the view in 4d from monocular egocentric video. InEuropean Conference on Computer Vision (ECCV) on Pith
  22. [42] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J. Davison. 2017. Scenenet rgb-d: Can 5m synthetic images beat generic imagenet pre-training on indoor segmentation? InProceedings of the IEEE International Conference on Computer Vision (ICCV) on Pith
  23. [43] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. 2024. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  24. [44] Yun Liu, Haolin Yang, Xu Si, Ling Liu, Zipeng Li, Yuxiang Zhang, Yebin Liu, and Li Yi. 2024. TACO: Bench- marking Generalizable Bimanual Tool-ACtion-Object Understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition on Pith
  25. [45] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. 2023. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception.Preprint, arXiv:2306.06362 on PithPith's reviewarXiv
  26. [46] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. 2024. Reconstructing hands in 3d with transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  27. [47] Toby Perrett, Ahmad Darkhalil, Saptarshi Sinha, Omar Emara, Sam Pollard, Kranti Parida, Kaiting Liu, Prajwal Gatti, Siddhant Bansal, Kevin Flanagan, Jacob Chalk, Zhifan Zhu, Rhodri Guerrier, Fahd Abdelazim, Bin Zhu, Davide Moltisanti, Michael Wray, Hazel Doughty, and Dima Dame... on Pith
  28. [48] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. 2024. Unidepth: Universal monocular metric depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  29. [49] Edwin Olson. 2011. Apriltag: A robust and flexible visual fiducial system. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3400–3407 on Pith
  30. [50] Ryan Punamiya, Simar Kareer, Zeyi Liu, Josh Citron, Ri-Zhao Qiu, Xiongyi Cai, Alexey Gavryushin, Jiaqi Chen, Davide Liconti, Lawrence Y. Zhu, Patcharapong Aphiwetsa, Baoyu Li, Aniketh Cheluva, Pranav Kuppili, Yangcen Liu, Dhruv Patel, Aidan Gao, Hye-Young Chung, Ryan Co, and 2... on PithPith's reviewarXiv
  31. [51] Patrick Rim, Kevin Harris, Braden Copple, Shangchen Han, Xu Xie, Ivan Shugurov, Sizhe An, He Wen, Alex Wong, Tomas Hodan, and Kun He. 2026. SHOW3D: Capturing scenes of 3d hands and objects in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re... on Pith
  32. [52] Javier Romero, Dimitrios Tzionas, and Michael J. Black. 2017. Embodied hands: Modeling and capturing hands and bodies together.ACM T ransactions on Graphics, (Proc. SIGGRAPH Asia), 36(6) on Pith
  33. [53] Ropedia. 2026. Xperience-10M: A large-scale egocentric multimodal dataset with structured 3D/4D annotations. Hugging Face. Dataset on Pith
  34. [54] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. 2025. WiLoR: End-to-end 3d hand localization and reconstruction in-the-wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 14 on Pith
  35. [55] Oriane Siméoni, Huy V . Vo, 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, C... on PithPith's reviewarXiv
  36. [56] Edgar Sucar, Eldar Insafutdinov, Zihang Lai, and Andrea Vedaldi. 2026. V-DPM: 4d video reconstruction with dynamic point maps. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  37. [57] Edgar Sucar, Zihang Lai, Eldar Insafutdinov, and Andrea Vedaldi. 2025. Dynamic point maps: A versatile representation for dynamic 3d reconstruction.Preprint, arXiv:2503.16318 on PithPith's reviewarXiv
  38. [58] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. 2025. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition on Pith
  39. [59] Ropedia R&D Team. 2026. HOMIE Gen2: Scale Real-World Data for Physical AI. https://ropedia.com/ homie. Accessed 2026-09-19 on Pith
  40. [60] Jikai Wang, Qifan Zhang, Yu-Wei Chao, Bowen Wen, Xiaohu Guo, and Yu Xiang. 2025. HO-cap: A capture system and dataset for 3d reconstruction and pose tracking of hand-object interaction. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and ... on Pith
  41. [61] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. 2025. Continuous 3d perception model with persistent state. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  42. [62] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. 2024. Dust3r: Geometric 3d vision made easy. InCVPR on Pith
  43. [63] Zikai Wang, Zhilu Zhang, Yiqing Wang, Hui Li, and Wangmeng Zuo. 2026. Arthoi: Taming foundation models for monocular 4d reconstruction of hand-articulated-object interactions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition on Pith
  44. [64] Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. 2026. VGGT- Ω. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  45. [65] Mingxi Xu, Bowen Duan, Yi Gu, Zhengyang Shen, Renjing Xu, and Yutao Yue. 2026. HandFlow: Fully Generative 4D Hand Recovery with Flow Matching.arXiv preprint arXiv:2607.11221 on PithPith's reviewarXiv
  46. [66] Yufei Ye, Jiaman Li, Ryan Rong, and C. Karen Liu. 2026. Whole: World-grounded hand-object lifted from egocentric videos.CVPR Findings on Pith
  47. [67] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. 2023. Metric3D: Towards zero-shot metric 3D prediction from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision on Pith
  48. [68] Zhao-Heng Yin, Changhao Wang, Luis Pineda, Francois Hogan, Krishna Bodduluri, Akash Sharma, Patrick Lan- caster, Ishita Prasad, Mrinal Kalakrishnan, Jitendra Malik, Mike Lambeta, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. 2025. Dexteritygen: Foundation controller for unpr... on PithPith's reviewarXiv
  49. [69] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. 2024. FoundationPose: Unified 6d pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition on Pith
  50. [70] Chuhan Zhang, Guillaume Le Moing, Skanda Koppula, Ignacio Rocco, Liliane Momeni, Junyu Xie, Shuyang Sun, Rahul Sukthankar, Joëlle K. Barral, Raia Hadsell, Zoubin Ghahramani, Andrew Zisserman, Junlin Zhang, and Mehdi S. M. Sajjadi. 2026. Efficiently reconstructing dynamic scene... on Pith
  51. [71] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandos Alexandros Potamias. 2025. Hawor: World-space hand motion reconstruction from egocentric videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  52. [72] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. 2025. MonST3r: A simple approach for estimating geometry in the presence of motion. InThe Thirteenth International Conference on Learning Representations on Pith
  53. [73] Zijie Zhu, Weiren Cai, Yizhou Wang, Zhenjie Yang, Yide Liu, Jiahao Chen, and Guanqi He. 2026. Mint: A unified model for world-space camera and hand motion estimation from scalable egocentric pipeline supervision.Preprint, arXiv:2609.04958. 16 Appendix Contents A Extended Relat... on PithPith's reviewarXiv
  54. [74] Zhengdi Yu, Stefanos Zafeiriou, and Tolga Birdal. 2025. Dyn-hamr: Recovering 4d interacting hand motion from a dynamic camera. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) on Pith
  55. [2022] Ego4d: Around the world in 3,000 hours of egocentric video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995–19012 on Pith
  56. [2023] Dynamicstereo: Consistent dynamic depth from stereo videos.CVPR on Pith
  57. [2025] Depth Anything 3: Recovering the Visual Space from Any Views.arXiv preprint arXiv:2511.10647v1. https://depth-anything-3.github.io/ on PithPith's reviewarXiv
  58. [2026] Ace-data-0: Human-centric ambient capture as embodied data engine.arXiv preprint arXiv:2607.28625 on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment