{"id":"4309c8a7-2f68-4d64-b21a-c22070fa6f0c","arxiv_id":"2505.11868","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MonoMobility recovers motion parts and their 3D axes from a monocular video by initializing with depth, flow, and point-cloud registration, then refining via 2D Gaussian splatting optimization.","lead":"This paper presents a method that takes a single ordinary video of an articulated object, such as a drawer or swivel chair, and works out which parts move and how they move in 3D, without needing a training dataset for the task. The significance is that robots could use such a capability to understand everyday objects from a single camera view, replacing setups that require multiple cameras or precise labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on monocular depth/flow quality, and the synthetic evaluation never isolates this dependency, so the reported 1.262° mean AE may not transfer to real videos.","rationale":"After reading the full manuscript, the central claim is an empirical one: zero-shot parsing to low error from monocular video. The method's own ablations show that the initialization is critical, and the initialization is derived from monocular depth and optical flow. The conclusion concedes these can fail. The quantitative evaluation is entirely on self-created synthetic scenes with Blender renders, where depth and flow are much cleaner than on real phone videos; real-world results are qualitative only. Therefore the most load-bearing uncertainty is whether the 1.262° mean AE reflects the proposed optimization or the favorable preprocessing conditions. The proposed test, replacing estimated depth/flow with ground-truth preprocessing, directly isolates this dependency. This is a stronger and more actionable version of the reader's identified weakest assumption. I do not see an internal inconsistency in the optimization; the main risk is external validity. Hence the verdict remains CONDITIONAL pending this test.","tokens_in":12806,"tokens_out":6634,"duration_ms":75853,"concrete_test":"On the authors' Blender dataset, render the same 15 scenes and replace the estimated depth and optical flow with Blender ground-truth depth/flow and known camera poses during preprocessing; re-run the full pipeline and compare AE, TA, and PE against Table 1. If the ground-truth-preprocessing result is substantially better than the estimated-preprocessing result (e.g., AE drops below 0.5°), the off-the-shelf estimates are the bottleneck and the abstract's 'precisely' claim is not robust to realistic depth/flow noise. If the results match, the concern is resolved. Report per-scene errors and at least 3 runs to assess stability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract promises precise parsing from a monocular video alone, but the coarse stage (Sec. 4.2–4.3) and the motion loss (Eq. 7) are built entirely on point clouds unprojected from off-the-shelf monocular depth, camera poses from MASt3R, and optical-flow segmentation. Monocular depth estimators output scale/shift-inconsistent depth across frames, so the 'largest-motion' pair used for axis initialization and the P_A/P_B supervision in Eq. 7 can carry a systematic bias that is not corrected by the rendering and normal losses. The paper acknowledges this in the conclusion: 'our method relies on offline depth estimation and optical flow estimation methods... they may perform poorly in certain special scenarios, leading to failures in the initial parsing results.' However, Table 1 is generated on Blender synthetic scenes where depth and flow are near-perfect; it therefore does not substantiate the unconditional central claim. The ablation in Table 2 shows that replacing the axis initialization with random initialization degrades TA from 1.0 to 0.667 and AE from 1.143° to 10.235°, so the method's accuracy is highly sensitive to the very initialization that depends on monocular depth/flow quality. Without quantifying this dependency, the headline result is conditional on unstated sensor and preprocessing assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MonoMobility, a zero-shot framework that takes a monocular video of an articulated object and outputs motion parts together with their motion attributes (axis direction, axis position, and motion type: translation, rotation, or rotation+translation). The method first uses off-the-shelf monocular depth estimation, optical flow, and MASt3R to build per-frame point clouds and segment candidate motion parts. It initializes a 2D Gaussian splatting scene from the first-frame point cloud, estimates an initial motion axis from the largest-motion pair of point clouds, and then optimizes per scene with rendering, normal, and motion losses. Motion parts are pruned and motion types are classified after optimization. Quantitative evaluation is carried out on a self-created Blender synthetic dataset (15 scenes) with comparisons against PARIS-scene, PARIS-obj, and DGMarbles*, and qualitative results are shown on 11 real-world videos. The reported mean axis-angle error is 1.262 degrees on the synthetic benchmark, with ablations showing the importance of axis initialization and of each loss term.","tokens_in":13075,"tokens_out":5458,"duration_ms":53560,"significance":"If the reported results transfer to real monocular videos, the work would be a useful step toward annotation-free articulated-object analysis for embodied agents, avoiding the dense multi-view or part-level supervision required by prior methods. The paper's strengths are its fully self-supervised optimization objective, explicit integration of motion attributes with 2D Gaussian splatting, and clean ablations showing that the axis initialization, rendering loss, normal loss, and motion loss each contribute to the final accuracy. The synthetic quantitative results are encouraging. However, the quantitative evidence is limited to a small synthetic dataset, the real-world evaluation is qualitative only, no error bars or multiple runs are reported, and no code is released; these gaps prevent the unconditional zero-shot claim from being fully supported.","major_comments":[{"comment":"The quantitative evaluation is run exclusively on synthetic Blender scenes, but the paper does not state whether the depth maps and optical flow used in Section 4.2 are rendered ground truth or outputs of the off-the-shelf estimators [12,20,21,49]. If they are ground truth, Table 1 measures only the optimization stage, not the full monocular-video pipeline; if they are estimated, the near-perfect synthetic geometry still gives no evidence about robustness to real depth/flow errors. Because the abstract claims precise parsing \"only using a monocular video\" and the conclusion explicitly acknowledges failure when depth/flow are poor, the paper needs a controlled study (e.g., perturbing depth/flow quality or evaluating on real videos with pseudo ground truth) that reports how AE, PE, and TA degrade.","section":"Section 5.1 and Table 1"},{"comment":"The motion loss uses point clouds P_A and P_B that are extracted from the same depth maps and segmentation used to initialize the motion-part Gaussians and the axis parameters. Consequently, the motion loss can reinforce the initialization rather than independently correct it; Table 2 shows that removing the axis initialization raises AE from 1.143 degrees to 10.235 degrees, underscoring the initial axis's influence. This is not fatal because the rendering and normal losses provide independent photometric and geometric supervision, but the paper should quantify how much of the final axis correction comes from the motion loss versus the rendering/normal losses, for example by ablating Eq. (7) under corrupted or random initialization.","section":"Eq. (7) and Section 4.4"},{"comment":"All quantitative tables report point estimates from a single optimization run with no standard deviations or number of seeds. Given the stochastic frame-pair sampling in Section 4.4 and the small number of synthetic scenes (15), the headline mean AE of 1.262 degrees may not be stable; the paper should report multiple runs with variance and per-category uncertainty. In addition, the thresholds in Eq. (4) (theta_min = 10 degrees, alpha_min = 0.1*radius, phi_min = 0.05*pi, iterjudge = 2000) are hand-set, and a sensitivity analysis is needed to support the claim of annotation-free operation.","section":"Tables 1–3"}],"minor_comments":[{"comment":"The expression \"qk = P k 1 / |P k 1|\" is ambiguous; it should be written as the centroid or mean of the first-frame point cloud, not as the point cloud divided by its cardinality.","section":"Section 4.3"},{"comment":"The entry \"1.68 1\" in the AE row for the Fridge*3 category appears to be a formatting error (likely 1.681); please fix the typesetting.","section":"Table 1"},{"comment":"The transformation matrix in the motion loss is written as \"Mat k A→B\" without a superscript in Eq. (7), while the text defines it as \"Mat k A→B = Mat k B Mat k A^−1\"; unify the notation and define all symbols consistently.","section":"Eq. (7)"},{"comment":"DGMarbles* uses ground-truth segmentation to partition the exported Gaussian point clouds, while MonoMobility uses estimated segmentation; the paper should state this asymmetry explicitly, since it makes the baseline comparison favorable to DGMarbles* in terms of part identification while still showing worse axis accuracy.","section":"Section 5.3"},{"comment":"The real-world evaluation is qualitative only; adding failure cases and a discussion of when the off-the-shelf depth/flow preprocessing fails would materially strengthen the robustness claims.","section":"Section 5.5 and Figure 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a genuine pipeline contribution, not a breakthrough. It combines existing monocular depth/flow/segmentation tools with a 2D Gaussian splatting optimization that treats motion parts as rigid groups. The strong numbers on synthetic scenes are real, but they are measured under near-perfect depth conditions, so treat the headline accuracy as conditional.\n\nWhat's new: the motion-part Gaussian representation and the three-loss objective are new as a combination. The idea of refining an initial coarse axis estimate with self-supervised rendering, normal, and motion consistency losses is sensible, and the ablations actually show each loss pulling its weight. The authors are also honest in the conclusion about relying on off-the-shelf depth and flow, and they flag the per-scene optimization cost.\n\nWhere it's soft: the quantitative evaluation is on a self-created Blender dataset. Baselines are adapted, which is fair, but no error bars are given. The real-world results are qualitative only, and the paper does not investigate how performance degrades as depth/flow quality worsens. The stress-test concern is legitimate: the initialization that seems to matter most (Table 2) is built from the same point clouds that supervise the motion loss, so a biased depth estimate could corrupt both. Rendering and normal losses provide independent signal, but the synthetic setting gives them near-perfect geometry to work with. The 1.262° mean AE is therefore best read as an upper-bound estimate for well-behaved videos, not a guarantee for arbitrary monocular input.\n\nDo I have a stronger objection? Not really. The method is honestly positioned as a pipeline, the limitations are stated, and the novelty is incremental but real. The biggest fix would be to add real-world quantitative data or a sensitivity analysis over depth/flow quality.\n\nBottom line: a solid, citable paper for embodied-intelligence folks working on articulated objects. Send it to review; it will need revision, but the core idea is worth engaging with.","headline":"Genuine pipeline contribution with a real conditional: the strong synthetic numbers assume near-perfect depth/flow, so the zero-shot claim is softer than the abstract suggests.","tokens_in":13568,"tokens_out":1949,"would_cite":true,"duration_ms":19583,"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":"A monocular video alone can parse the moving parts of articulated objects and recover their 3D motion axes and motion types, without annotated training data.","keywords":["articulated objects","zero-shot motion analysis","monocular video","2D Gaussian splatting","motion axis estimation","optical flow segmentation","dynamic scene optimization","motion type classification"],"falsifier":"Record or render a monocular video of an articulated object in a setting known to break monocular depth or optical flow, such as a specular, textureless rotating part or a slow-moving part whose optical-flow segmentation merges it with the static background, and check whether the pipeline still finds the moving part and its axis. Because the paper states that no Gaussians are added or removed during optimization and that only pruning occurs, any moving part missed by the initial optical-flow segmentation is unrecoverable; a clearly visible small-amplitude motion that the initializer misclassifies as static would refute the universality of the zero-shot claim.","tokens_in":12626,"feed_emoji":"🎥","tokens_out":5336,"duration_ms":49704,"temperature":0.7,"pith_summary":"This paper tries to show that the motion parts and motion attributes of articulated objects can be extracted from a single monocular video, with no annotated training data. It combines off-the-shelf depth and optical flow estimates to initialize a 2D Gaussian splatting scene, then optimizes the scene with rendering, normal, and motion losses to refine the parts and their axes. If correct, robots with ordinary RGB cameras could analyze interactions with doors, drawers, and staplers in unfamiliar scenes without retraining. On a synthetic benchmark the method reports a mean axis-angle error around 1.26 degrees and correctly classifies rotation, translation, and combined motions.","feed_headline":"Monocular video alone yields 3D motion axes for articulated parts","feed_subtitle":"No annotations, no multi-view rigs: mean axis error about 1.3 degrees on synthetic scenes.","key_machinery":"The load-bearing mechanism is a 2D Gaussian splatting scene, a differentiable surface representation made of flat Gaussian primitives rendered by alpha-blending, in which each motion part is a rigidly transformed group of Gaussians governed by a shared motion attribute consisting of axis direction, axis position, and per-frame translation and rotation. A coarse initial parse, built from depth-derived point clouds, optical-flow segmentation, and point-cloud registration between the two frames with largest motion, provides the starting axis and part labels. An end-to-end optimization then refines all parameters against three losses: a rendering loss for visual consistency, a normal loss for geometric consistency, and a motion loss that keeps each Gaussian group's transformation consistent with the point-cloud motion between sampled frame pairs. The motion-type decision (translation, rotation, or both) and the pruning of spurious parts come from thresholding the accumulated translation and rotation after a fixed number of iterations.","core_discovery":"The central claim is that 3D mobility analysis, identifying which parts move and how, can be posed as self-supervised dynamic scene optimization rather than a supervised recognition problem. Starting from monocular depth, optical flow, and camera poses supplied by pretrained estimators, the method segments candidate moving parts, initializes each as a group of 2D Gaussians with an assumed rotation-plus-translation axis, and then jointly optimizes the scene and the motion parameters so rendered frames, surface normals, and point-cloud motion all agree. The optimization both prunes false motion parts and refines axis direction and position, after which motion type is read off from the total translation and rotation magnitudes. The paper reports that this yields lower mean angular error than compared baselines, with mean error 1.262 degrees on its synthetic scenes, and plausible qualitative results on phone-captured real videos.","pith_inferences":["Editorial inference: if the underlying depth and flow estimators improve, the same coarse-to-fine optimization could extend to non-rigid articulated motion by replacing rigid part transforms with per-part warp fields.","Editorial inference: the framework could act as a pseudo-label generator, producing axis and part annotations on unlabeled videos that could train a fast feed-forward predictor, moving the per-scene optimization cost offline.","Editorial inference: the motion-loss idea, matching Gaussian-group transformations to point-cloud transformations, is a modular constraint that could be adapted to other dynamic scene representations, such as dynamic meshes or implicit surfaces."],"forward_implications":["Robots equipped with ordinary monocular RGB cameras could parse articulated-object motions in unfamiliar scenes without per-object annotations or multi-view rigs.","The method covers pure translation, pure rotation, and combined rotation-plus-translation, so it generalizes beyond simple hinges and sliders to motions like the helical path of a lift chair.","Because the optimization is per-scene and self-supervised, any video of a moving articulated object, including casual phone recordings, can serve as input, though at the cost of per-scene optimization time.","Ablations indicate the optimization contributes mainly to axis and type refinement: motion-type accuracy rises from 0.667 to 1.0 when axis initialization is added, while the part-segmentation IoU stays at 0.956."],"supporting_citations":[{"why":"Provides the 2D Gaussian splatting scene representation and the rendering loss used for optimization.","marker":"[14]"},{"why":"Estimates camera poses from the monocular video, needed to lift depth maps into a common world point cloud.","marker":"[21]"},{"why":"One of the off-the-shelf monocular depth estimators whose outputs seed the scene geometry.","marker":"[12]"},{"why":"Another off-the-shelf monocular depth estimator used to build the initial point clouds.","marker":"[49]"},{"why":"Supplies optical flow used to track pixel trajectories and segment candidate motion parts.","marker":"[40]"},{"why":"Segments the optical flow into instance-level motion masks that define the initial motion parts.","marker":"[47]"},{"why":"A multi-view articulated-object baseline that the paper adapts and compares against for axis accuracy.","marker":"[25]"},{"why":"A monocular dynamic-scene reconstruction baseline that, with post-processing, serves as the main comparison for axis estimation.","marker":"[37]"}],"fun_headline_variants":["Zero-shot 3D mobility from monocular video alone","No annotations: video alone reveals 3D motion axes","Self-supervised 3D motion parsing from a single video","Monocular video yields 3D articulated motion without labels","One video, no training: 3D mobility analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the off-the-shelf depth and optical flow estimators provide point clouds accurate enough that the initial axis estimate and the motion loss are both trustworthy; if those pretrained models fail on a scene, the initial parsing breaks and the optimization cannot recover the true motion parts.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot 3D mobility from monocular video alone","No annotations: video alone reveals 3D motion axes","Self-supervised 3D motion parsing from a single video","Monocular video yields 3D articulated motion without labels","One video, no training: 3D mobility analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1380,"prompt_tokens":938,"completion_tokens":442,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":360}},"tokens_in":554,"tokens_out":442,"duration_ms":4119,"temperature":1.0,"reasoning_tokens":360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:45:29.367710+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record or render a monocular video of an articulated object in a setting known to break monocular depth or optical flow, such as a specular, textureless rotating part or a slow-moving part whose optical-flow segmentation merges it with the static background, and check whether the pipeline still finds the moving part and its axis. Because the paper states that no Gaussians are added or removed during optimization and that only pruning occurs, any moving part missed by the initial optical-flow segmentation is unrecoverable; a clearly visible small-amplitude motion that the initializer misclassifies as static would refute the universality of the zero-shot claim.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Provides the 2D Gaussian splatting scene representation and the rendering loss used for optimization."},{"cited_title":"Ground- ing image matching in 3d with mast3r","cited_arxiv_id":null,"evidence_quote":"Estimates camera poses from the monocular video, needed to lift depth maps into a common world point cloud."},{"cited_title":"Paris: Part-level reconstruction and motion analysis for articulated objects","cited_arxiv_id":null,"evidence_quote":"A multi-view articulated-object baseline that the paper adapts and compares against for axis accuracy."}],"review_version":1}