{"id":"0eae7c81-3d6a-439e-a509-6708e033a506","arxiv_id":"2512.07394","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A constrained optimisation-and-propagation method that enforces stable hand contact while an object is held improves 3D pose reconstruction of rigid objects across egocentric hand-interaction timelines.","lead":"This paper introduces ROHIT, a task for reconstructing a rigid object's 3D pose across a full hand-interaction timeline in egocentric video, from static to grasped to released. It proposes a stable-grasp constraint plus pose propagation (COP) and contributes annotated stable-grasp and timeline clips from HOT3D and EPIC-Kitchens.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"E_SG (Eq. 6) enforces pairwise distance invariance across all object-fingertip pairs, stronger than the paper's contact-area IoU definition (Eq. 1); rolling/sliding stable grasps would be mis-optimized, and SCA metrics partly reward the same objective.","rationale":"The reader's weakest_assumption is precisely the over-strong nature of E_SG: it assumes distance invariance across all frames, which is not implied by the contact-area IoU definition. My attack sharpens this by noting that on EPIC-HIT the SCA metrics are the same quantity E_SG optimizes, making the in-the-wild gains partly circular. However, this concern is not fatal: the HOT3D ADD results provide independent evidence that E_SG helps in at least some stable-grasp configurations, and the paper honestly discloses the hand-pose and boundary dependencies. The correct response is to retain the CONDITIONAL verdict, requiring the proposed diagnostic test and ideally a soft-constraint variant to establish generality. Since the reader already reached CONDITIONAL, my read does not change the verdict.","tokens_in":22881,"tokens_out":6361,"duration_ms":62144,"concrete_test":"Using HOT3D ground truth, compute for each stable-grasp segment the mean per-vertex coefficient of variation of d_oh (Eq. 7) across all object-fingertip pairs. Split segments by this variation and compare COP vs Dynamic ADD: if the gain disappears or becomes negative for the high-variation half, the distance-invariance assumption is load-bearing. Also re-run the EPIC-HIT comparison with an alternative E_SG that enforces only the contact-area IoU of Eq. (1) (e.g., penalize changes in the set of vertices with d_oh < ε). If the SCA improvements over Dynamic shrink or vanish, the reported in-the-wild gains partly reward the exact objective being optimized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the stable-grasp loss E_SG yields more accurate object poses. Eq. (6) minimizes the variation of d_oh = |v_o - v_h| between every object vertex and every fingertip vertex across all frame pairs, requiring the entire distance field to be time-invariant. This is much stronger than the paper's own formal definition of stable grasp (Eq. 1), which only requires IoU of the in-contact surface area > τ and permits rolling/sliding or changes in non-contact distances. On HOT3D, ground-truth masks and ADD evaluation limit the damage. But on EPIC-HIT—where there is no 3D ground truth—the reported SCA-IOU and SCA@0.8 metrics directly measure contact-area stability, exactly what E_SG is designed to enforce; the gains over Dynamic there are therefore partly circular. The supplementary boundary-noise ablation (Table 8) and acknowledged hand-pose failures (Fig. 8, Sec. G) show sensitivity to inputs but do not quantify the over-constraint. The HOT3D ADD results remain the only independent evidence, and they may transfer only to grasps where distance invariance approximately holds; the paper never measures how often that condition is satisfied in its own stable-grasp data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ROHIT, a task for reconstructing rigid object 6-DoF poses over a Hand Interaction Timeline (HIT) in egocentric video, and COP, an optimization framework with segment-specific constraints: Static segments fix the object in the world, Stable Grasp segments use a novel stable-grasp loss E_SG that penalizes variation of hand-object vertex distances across frames, and Unstable Contact segments use mask and physical-contact losses. COP also propagates optimized object-to-world poses across segment boundaries to initialize subsequent segments. The authors curate HOT3D-HIT (1,239 stable grasps, 113 HITs with 3D GT) and EPIC-HIT (2,431 stable grasps, 96 HITs with 2D masks), and report ADD improvements on stable grasps from 51.9% to 58.1% (HOT3D) and 56.0% to 65.1% (ARCTIC), with further gains from propagation, e.g., stable-grasp ADD 70.3% and Static ADD 79.2% on HOT3D-HIT.","tokens_in":23236,"tokens_out":8142,"duration_ms":79044,"significance":"If the reported gains are robust, the paper makes a useful contribution: it defines a new temporal task with explicit segment types, introduces two datasets, and shows that enforcing stable-contact constraints and propagating poses across segments improves object pose estimation. The strongest evidence is the GT-based ADD evaluation: the HOT3D-HIT and ARCTIC results are not built from the SCA metrics that E_SG targets, and the supplementary ARCTIC evaluation plus the boundary-noise ablation provide useful supporting analyses. However, the EPIC-HIT quantitative claims rely on SCA metrics that are closely aligned with the optimized loss, and the paper's own definition of stable grasp is weaker than the constraint imposed by E_SG; these issues need additional analysis before the central claims can be fully accepted.","major_comments":[{"comment":"E_SG enforces a much stronger condition than the paper's own stable-grasp definition. Eq. (1) only requires the IoU of the in-contact object area to exceed tau, permitting gradual rolling/sliding or changes in non-contact distances. Eq. (6) instead minimizes, for every object vertex and every fingertip vertex, the variation of d_oh over all frame pairs, i.e. it requires the full distance field to be time-invariant. Grasps that satisfy Eq. (1) but include slight rolling/sliding or soft-tissue deformation can then be optimized to a wrong pose. The GT ADD results on HOT3D/ARCTIC are not affected by this concern, but the paper does not measure how often the stronger condition actually holds in its own stable-grasp data. Please report the distribution of contact-area IoU vs. Eq. (6) distance variation under GT poses, and consider a relaxed loss that constrains only the nearest-contact vertice","section":"Sec. 4.2, Eq. (6)"},{"comment":"The SCA@0.8/0.6 and SCA-IOU metrics measure exactly the quantity that E_SG is designed to optimize: consistency of in-contact areas across frame pairs. The paper's own explanation in Sec. 4.2 that aligned distance fields imply aligned contact area makes the EPIC-HIT gains on SCA metrics partly circular. Since EPIC-HIT has no 3D GT, the only independent support comes from HOT3D-HIT and ARCTIC ADD. The abstract and Sec. 1 should not present EPIC-HIT SCA gains as direct evidence of pose accuracy. Please report the relationship between IOU and ADD, or add a no-GT proxy that is not the loss objective.","section":"Sec. 5.3, Table 3"},{"comment":"The hyperparameters lambda1 and lambda2 are selected by measuring ADD/SCA-ADD on HOT3D-HIT, which is the same set used to report the headline results in Tables 2 and 4. No validation split is described. This selection on the evaluation set can inflate reported improvements; the table shows, for instance, lambda1=10 improves SCA-ADD but hurts ADD relative to lambda1=1, so the choice is consequential. Please use a held-out validation split or report sensitivity to lambda over a range.","section":"Table 7"},{"comment":"The Unstable Contact improvement from 17.3 to 67.1 ADD is the largest component of the claimed up-to-24.5% HIT gain, but COP without propagation uses random initialization for Unstable Contact while COP with propagation initializes from the optimized previous segment, which in HOT3D is itself initialized from GT-derived poses. This comparison primarily measures the value of a good initialization rather than a temporal constraint. COP's propagation is a warm start, not a joint temporal optimization (Sec. 4.4). Please add an ablation that initializes both variants with the same initialization protocol, or add a temporal consistency term, to separate initialization effects from constrained propagation.","section":"Table 4 and Supplement Sec. E"}],"minor_comments":[{"comment":"The 'whiteboardmarker' row reports COP with SCA-IOU 0.0 but SCA-ADD 35.4; since both are averages over the same category, this inconsistency needs an explanation or correction.","section":"Table 2"},{"comment":"The definition of Unstable Contact uses IOU(Si,Sj)<=tau, but if Si or Sj is empty the IoU is undefined. Please clarify how empty contact sets are handled.","section":"Sec. 3.1, Eq. (1)"},{"comment":"The range '6.2-11.3%' is not tied to specific datasets or tables in the abstract; it appears to refer to HOT3D and ARCTIC ADD gains. Please state the source of each number.","section":"Abstract"},{"comment":"The EPIC-HIT row lists 79,736 frames for 96 HITs, while Supp. B.2 reports 319,661 segmentation masks for stable-grasp clips. Please clarify whether the table counts only HIT frames or includes stable-grasp clips.","section":"Table 1"},{"comment":"The scaling lambda_f = f * render_size needs units and a precise definition of f, since Eq. (5) is in pixels and Eq. (6) is in meters; this is needed for reproducibility.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a vision venue, and the GT-based ADD results are the main reason I am not recommending rejection. The revision should focus on strengthening the evidence around E_SG: quantify the gap between Eq. (1) and Eq. (6), de-emphasize the circular SCA-based claims for EPIC-HIT, and avoid hyperparameter selection on the test set. These are fixable with additional experiments and re-analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper introduces a new task (ROHIT) with a temporal segmentation of hand-object interaction into Static, Unstable Contact, and Stable Grasp, curates two annotated benchmarks, and proposes a constrained optimization plus propagation method (COP). The strongest evidence is on HOT3D, where COP improves ADD on stable grasps from 51.9 to 58.1 without propagation, and to 70.3 with propagation on full timelines. The ARCTIC results in the supplement are consistent. That is a genuine, independent signal.\n\nWhat is actually new: the HIT formulation is new, and the two datasets are a real artifact — HOT3D-HIT gives 3D ground truth, EPIC-HIT goes in-the-wild with manually annotated stable grasps and timelines. E_SG, the temporal stable-contact loss, is a reasonable idea. The propagation itself is an extension of HOMan rather than a breakthrough, but applying it across segment types is still a useful contribution.\n\nSoft spots, in order of severity. First, the stress-test is right: Eq. (6) enforces invariance of all object-to-fingertip distances across all frame pairs, which is stronger than Eq. (1)'s contact-area IoU. Rolling or sliding grasps, soft-tissue deformation, or hand-pose errors can violate that, and the paper acknowledges the latter (Fig. 8, Sec. G) but never measures how often the distance-invariance condition actually holds. Second, the SCA metrics on EPIC-HIT partly reward exactly what E_SG optimizes, so those gains are partially circular. The HOT3D ADD results are the only fully independent evidence, and they may overstate transfer to the wild. Third, the loss weights (Table 7) are selected on the evaluation data without a held-out split, and there are no error bars or significance tests. These are addressable rather than fatal.\n\nThe paper is honest about its limitations — it relies on known CAD models and accurate segment boundaries, and the boundary-noise ablation (Table 8) shows a real drop. The central claim, that enforcing stable contact and propagating poses across segments improves rigid object pose reconstruction in egocentric video, holds up on the GT data.\n\nWho it is for: people working on egocentric hand-object reconstruction and object tracking through occlusion. The benchmarks will probably get used. A serious referee should engage with it, not desk-reject. My recommendation to the editor and to you: send it to review, and push for three things in revision — release code and annotations, use a proper validation split for hyperparameters, and quantify how often the E_SG assumption actually holds in the stable-grasp data.","headline":"Solid task-definition and dataset paper; the GT-based ADD gains on HOT3D are real, but the EPIC SCA metrics are partly circular and the stable-grasp loss is stronger than the paper's own definition.","tokens_in":23753,"tokens_out":2227,"would_cite":true,"duration_ms":23235,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that enforcing stable-grasp contact invariance across frames and propagating poses across interaction segments reconstructs object pose in egocentric video better than per-frame fitting, improving accuracy by up to 24.5%.","keywords":["Hand Interaction Timeline","stable grasp","egocentric video","object pose reconstruction","constrained optimisation","pose propagation","hand-object interaction","6-DoF pose estimation"],"falsifier":"A direct check: take stable-grasp clips where the object is known to rotate slightly in the hand (rolling or sliding) and where ground-truth poses are available. If COP's ADD on those clips is no better than the Dynamic per-frame baseline, or if the E_SG loss is lower for an incorrect pose than for the ground-truth pose, the contact-invariance constraint is not selecting the right pose. More practically, an experiment with deliberately perturbed hand poses (e.g., adding noise to the estimated fingers) should show COP degrading to chance levels, confirming that the frame-to-frame vertex-distanc","tokens_in":22761,"feed_emoji":"🤲","tokens_out":7757,"duration_ms":70977,"temperature":0.7,"pith_summary":"This paper introduces ROHIT, the task of reconstructing a known rigid object's 3D pose along a Hand Interaction Timeline in egocentric video—the moments when the object sits static, is picked up, held in a stable grasp, used, and set down. The authors argue that per-frame fitting ignores the structure of these interactions, so they propose COP, a constrained optimization that treats each timeline segment with its own constraint: during a stable grasp the same hand and object vertices stay in contact, so the distances between fingertip vertices and object vertices should be invariant across frames. COP then propagates the reconstructed object-to-world pose from one segment to the next as initialization. On their curated HOT3D-HIT dataset with 3D ground truth, COP raises stable-grasp pose accuracy from 51.9% to 58.1% ADD over per-frame dynamic fitting, and on full timelines raises unstable-contact accuracy from 17.3% to 67.1%; on in-the-wild EPIC-HIT with only 2D masks, it shows consistent gains. The significance is a step toward reconstructing objects in unconstrained egocentric video without 3D supervision.","feed_headline":"Stable-grasp constraint lifts object pose accuracy by up to 24.5%","feed_subtitle":"By keeping fingertip-to-object distances constant in a stable grasp, COP reconstructs objects from 2D masks alone.","key_machinery":"The load-bearing mechanism is the Hand Interaction Timeline decomposition into Static, Unstable Contact, and Stable Grasp segments, paired with the stable-grasp loss E_SG—a differentiable term that penalizes, for all pairs of frames in a segment, the variation of the distance between each object vertex and each fingertip vertex. E_SG carries the stable-grasp assumption into a numerical objective, and the propagation step—using the object-to-world pose at the segment boundary to initialize the next segment—is what links the timeline together. The push/pull losses and mask rendering loss handle physical contact and 2D supervision.","core_discovery":"The central claim is that a rigid object's pose during a hand interaction can be recovered from egocentric video by explicitly modeling the three temporal regimes the object passes through. In a stable grasp, the contact area on the object remains the same even though finger articulations and object pose vary, so the paper introduces the stable-grasp loss E_SG: for every object vertex and every fingertip vertex, the distance between them must be equal in every pair of frames of the segment. Because the object is rigid, minimizing E_SG by adjusting its pose relative to the hand aligns the contact region over time without freezing the object's motion. Each segment is then optimized under its o","pith_inferences":["A natural extension is to make the hand pose itself part of the optimization under E_SG; the paper's stated failure mode—incorrect hand poses—might then be mitigated, since the same contact-invariance objective would also constrain the hand.","The timeline decomposition suggests a benchmark protocol for in-the-wild egocentric reconstruction: evaluate contact-area consistency (SCA) alongside 2D mask IoU, so methods that overfit to masks but break contact physics are penalized.","The propagation scheme implies that errors can accumulate across long timelines; combining COP with an automatic segment localizer (e.g., a vision-language model) would be a direct next step, and the boundary-noise ablation indicates where such a model would help most.","One could test the stable-grasp assumption more directly by measuring whether E_SG is minimized near ground-truth poses on clips with deliberate micro-rolling; if not, the loss would need to be softened to tolerate slow contact drift."],"forward_implications":["Stable-grasp reconstruction improves by 6.2–11.3% over dynamic per-frame fitting, and full-HIT unstable-contact segments, the hardest part of an interaction, improve by up to 24.5% when poses are propagated.","On HOT3D-HIT, full-timeline ADD reaches 70.3% for stable grasps, 79.2% for static segments, and 67.1% for unstable contact, so the benefit is not limited to the in-hand portion.","Accuracy degrades gracefully under noisy segment boundaries: adding 10–30 frames of noise to boundaries lowers ADD from 70.0 to 60.0, still far above the 15.0 of the smoothing baseline, so the method does not require perfect annotations.","Because the optimization is training-free and only needs a category CAD model and 2D masks, it transfers to in-the-wild footage where 3D ground truth does not exist, and it also improves stable grasps in the ARCTIC dataset (ADD 56.0 to 65.1).","The proposed SCA-ADD and SCA-IOU metrics, which measure whether the reconstructed pose keeps a stable contact area, provide a way to evaluate contact consistency rather than only final pose."],"fun_headline_variants":["Stable-grasp constraint boosts object reconstruction by up to 24.5%","Modeling hand interaction timelines reconstructs objects in egocentric video","COP uses stable grasps to improve object reconstruction from video","Reconstructing objects from hand interactions: stable grasp helps","Egocentric object reconstruction boosted by stable-grasp modeling"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central assumption is that during a stable grasp the distance between every object vertex and every fingertip vertex is exactly the same in every frame—stronger than just 'the contact area overlaps,' and violated when the grasp rolls or slides slightly or when the hand pose is estimated incorrectly.","fun_headline_variants_meta":{"raw":{"variants":["Stable-grasp constraint boosts object reconstruction by up to 24.5%","Modeling hand interaction timelines reconstructs objects in egocentric video","COP uses stable grasps to improve object reconstruction from video","Reconstructing objects from hand interactions: stable grasp helps","Egocentric object reconstruction boosted by stable-grasp modeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000981,"raw_usage":{"total_tokens":4043,"prompt_tokens":828,"completion_tokens":3215,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":3127}},"tokens_in":572,"tokens_out":3215,"duration_ms":22311,"temperature":1.0,"reasoning_tokens":3127,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:56:31.969958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: take stable-grasp clips where the object is known to rotate slightly in the hand (rolling or sliding) and where ground-truth poses are available. If COP's ADD on those clips is no better than the Dynamic per-frame baseline, or if the E_SG loss is lower for an incorrect pose than for the ground-truth pose, the contact-invariance constraint is not selecting the right pose. More practically, an experiment with deliberately perturbed hand poses (e.g., adding noise to the estimated fingers) should show COP degrading to chance levels, confirming that the frame-to-frame vertex-distanc","supporting_citations":[],"review_version":1}