{"id":"bfb25fce-1278-46a7-947a-3758cbe08e72","arxiv_id":"2507.00339","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MOVi-MC-AC is a 2,041-scene multi-camera synthetic video dataset providing amodal masks, amodal content, depth, and consistent object IDs across six cameras, the first with ground-truth amodal content labels.","lead":"MOVi-MC-AC is a new synthetic dataset of 2,041 multi-camera video scenes showing household objects, with about 5.8 million labeled object instances. It is the first dataset to offer ground-truth amodal content, the hidden appearance of objects behind occlusions, to train computer vision models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth amodal content may be unverified: the dataset generation pipeline for \"unoccluded object appearance\" is not documented, and no code or consistency checks accompany the release, so the central claim \"first ground-truth amodal content dataset\" rests on an untested assumption.","rationale":"The reader's verdict is CONDITIONAL with the weakest assumption being that the ground-truth amodal content and multi-camera IDs are accurate and aligned. My review finds this to be the most load-bearing concern as well: the paper's principal novelty is precisely the provision of this ground-truth data, and the manuscript contains no verification of the annotations, no generation code, and no consistency checks. I do not find a more fundamental flaw in the paper's argument itself. The dataset is plausibly correct given that it is built from Kubric, a well-established simulator, and the authors do provide a substantial release. However, because the central claim depends entirely on annotation correctness, and that correctness is currently unverified, the CONDITIONAL verdict is appropriate. My concrete test directly validates the central claim on a sample, which would settle the concern. I do not see grounds for REJECT, because the concern is addressable and the dataset appears as described based on available text; UNCHANGED is the right recommendation, with the caveat that if the sample validation or metadata inspection reveals errors, the verdict would move toward REJECT.","tokens_in":7825,"tokens_out":1553,"duration_ms":15110,"concrete_test":"Inspect the HuggingFace repository for generation/validation metadata. If missing, independently validate a random sample of 20 scenes by recomputing occluded regions: for each sampled object, take the modal mask, amodal mask, and amodal content, and compare the amodal content pixel colors against a fresh Kubric render of the same scene with all other objects removed, using the same camera. If the amodal content differs materially or the identity association across cameras is inconsistent in any sampled scene, the central claim fails. Alternatively, publish code and report per-scene cross-camera ID agreement (should be 100%).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that MOVi-MC-AC is the first dataset to provide ground-truth amodal content, with ~5.8M instances across 2,041 multi-camera scenes (Table 1, Section 1). For this claim to hold, each object's amodal content must be exactly and reproducibly rendered: the RGB of every object pixel, including parts occluded by other objects, must come from the object's own unoccluded appearance, and object instance IDs must be consistent across all six cameras. The paper states only that amodal content is \"the ground-truth unobscured view of the object generated\" (Figure 2c) and that scene-object descriptors associate objects to scenes. It does not specify which Kubric renderer produced the amodal content, how the occluder/occludee separation was performed, how amodal depth was rendered, or how per-camera object IDs were assigned and verified. No cross-camera identity agreement statistics, no occlusion-consistency checks, and no generation code are provided, despite the full dataset being released on HuggingFace. The release appears to be raw data without annotation metadata or validation scripts. Consequently, the strongest claim, that these are true ground-truth amodal content labels, is currently unsubstantiated; if the amodal content is actually cut-and-paste composited from the visible parts, or if object identities drift across cameras, the dataset's central contribution collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MOVi-MC-AC, a synthetic multi-camera video dataset generated with Kubric. The dataset is claimed to be the first to provide ground-truth amodal content labels and the largest amodal segmentation dataset to date, with 2,041 scenes, 293,904 scene images, and 5,899,104 object instances across six cameras per scene. The paper presents dataset statistics in Table 1, describes scene-level and object-level annotations, and proposes evaluation metrics for amodal segmentation and amodal content completion adapted from existing mIoU, PSNR, LPIPS, and SSIM measures. The full dataset is released on HuggingFace. The paper does not include experiments, baselines, or code for regenerating or validating the annotations.","tokens_in":8218,"tokens_out":2669,"duration_ms":31081,"significance":"If the annotation claims are verified, this dataset would be a valuable resource: it is the first to offer ground-truth amodal content for occluded objects, and it provides consistent object IDs across multiple camera views, enabling new research on multi-camera amodal segmentation, content completion, and view-invariant object representation. The release on HuggingFace and the use of standard metrics are practical strengths. However, because the central claims depend entirely on the correctness of unverified synthetic annotations, the significance is conditional on the dataset generation and validation details being made available and confirmed.","major_comments":[{"comment":"The central claim that MOVi-MC-AC provides ground-truth amodal content is not supported by a description of how that content is generated. The text says only that amodal content is 'the unoccluded amodal RGB content' and that amodal depth is provided, but it does not specify the rendering pipeline: whether each object is rendered with other objects hidden, how occluder/occludee interactions are handled, how amodal depth is produced, or how the amodal content is aligned with the modal masks. Without this specification, the claim of ground-truth amodal content is unverifiable. Please document the exact rendering steps and provide visual and quantitative validation (e.g., checks that occluded pixels in the amodal content match the object's appearance in frames where it is unoccluded, or from a camera with an unobstructed view).","section":"Section 3, Annotations"},{"comment":"The dataset claims consistent object IDs across all six cameras, but no statistics or verification are reported. The paper should report the fraction of object instances whose IDs are consistent across all cameras, the rate of missing detections per camera, the number of fragmented or switched tracks, and agreement between scene-object descriptors and per-camera instance masks. Without such numbers, the multi-camera re-identification claim is not established.","section":"Table 1 and Section 3, scene-object descriptors"},{"comment":"The paper does not state that generation code, validation scripts, or annotation metadata are released. Since the contribution is a dataset, reproducibility requires either releasing the code used to generate the annotations or providing detailed pseudocode plus validation scripts that verify the amodal content, amodal masks, and cross-camera object ID consistency. At present, users cannot distinguish between correct annotations and plausible-looking but incorrect ones.","section":"Section 3, Data release"}],"minor_comments":[{"comment":"The definitions of mIoU and mIoUocc use symbols Mi, Ai, and Ai without clearly specifying that Mi is the modal mask, Ai is the ground-truth amodal mask, and Ai-hat is the predicted amodal mask; please define all symbols explicitly.","section":"Section 4.1"},{"comment":"Equation (3) for SSIM repeats the sentence 'x and y are the two image patches being compared' and the definitions of variables are duplicated; please clean up the exposition.","section":"Section 4.2"},{"comment":"There is a typo: 'scebe-object descriptors' should be 'scene-object descriptors'.","section":"Section 3, Annotations"},{"comment":"The number of instances is given as '5.8 million' in the abstract and '5,899,104' in Table 1; the consistency is fine, but the paper should clarify that 5.8 million is rounded.","section":"Abstract and Section 3"},{"comment":"The captions are long and nearly identical to each other; shorten them and ensure each caption identifies the specific content shown in the figure.","section":"Figure 1 and Figure 2"},{"comment":"The occluded-region metrics are defined only for images; for video content completion, clarity is needed on whether the metrics are computed per frame and then averaged over frames or computed on concatenated volumes.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a dataset release with no experiments, which is acceptable for some venues, but the central claims hinge on annotation correctness and cross-camera consistency. The authors should be asked to provide the generation pipeline, validation statistics, and either code or detailed pseudocode. I would also check whether the novelty relative to existing MOVi datasets is clearly distinguished, since the contribution appears to be annotation layers added to a known simulator rather than a new simulation engine."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe quick take: this is a dataset paper, and the dataset is a real contribution. MOVi-MC-AC gives you multi-camera video with amodal masks, amodal RGB content, amodal depth, and consistent object IDs across six cameras -- roughly 5.8 million instances and 2,041 scenes. That combination is new relative to SAIL-VOS, COCOA, D2S, and DYCE, and it is genuinely useful for the amodal segmentation and multi-camera tracking communities. The scale claims are internally consistent with the Kubric-based generation described in Section 3.\n\nCredit where due: the authors built on Kubric, which is a solid choice for this kind of synthetic data. The scene-object descriptors are a sensible addition for re-id tasks. The metrics section is standard but clearly defined. The full dataset is posted on HuggingFace, so the community can inspect it.\n\nThe soft spot is real and it is the load-bearing one. The paper states that the amodal content is the 'ground-truth unobscured view of the object generated,' but it never documents how that rendering is produced -- whether each object is rendered separately in an empty scene, whether occluders are excluded, or whether the 'amodal' content is actually composited from visible parts. It also reports no cross-camera identity agreement statistics and releases no generation code or validation scripts. If the amodal content is cut-and-paste from visible regions, or if object IDs drift between cameras, the central claims fail. These are addressable issues -- a code release and a few consistency checks would settle them -- but as submitted the verifiability is missing.\n\nThere are no experiments, which is fine for a dataset release, though the paper could have included at least one baseline to demonstrate that the annotations are trainable. The self-citations are not a problem.\n\nAll in all: the paper deserves serious peer review, but the referee should require the code, the rendering details, and verification statistics before accepting. For anyone working on amodal completion or multi-view tracking, this is a potentially valuable benchmark; the version of record should be one the community can trust.\n\nRecommendation: send to review, with conditional acceptance hinging on code and annotation verification.","headline":"A genuinely new multi-camera amodal dataset whose ground-truth claims are unverified as submitted; worth reviewing, but only after code and verification are provided.","tokens_in":8634,"tokens_out":2734,"would_cite":true,"duration_ms":28378,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new synthetic dataset, MOVi-MC-AC, provides the first ground-truth amodal content labels for roughly 5.8 million object instances in 2,041 multi-camera scenes, enabling direct training and evaluation of amodal content completion and…","keywords":["amodal segmentation","amodal content completion","multi-camera video","synthetic dataset","object re-identification","video object tracking","view-invariant representation","ground-truth amodal content"],"falsifier":"A reader could take a random sample of occluded instances, independently re-render each object's unoccluded appearance from its 3D model, and compare pixel-by-pixel with the provided amodal content; any systematic misalignment, or any object whose ID is not present in all six cameras' annotation files, would falsify the dataset's central claims.","tokens_in":7609,"feed_emoji":"🎥","tokens_out":7228,"duration_ms":74536,"temperature":0.7,"pith_summary":"The paper's claim is that computer vision's ability to reason about occluded objects has been limited by the absence of a large dataset with ground-truth amodal content — the full, unobscured appearance of every object — and it introduces MOVi-MC-AC to fill that gap. MOVi-MC-AC provides roughly 5.8 million object instances across 2,041 synthetic multi-camera scenes, with six cameras per scene and consistent object IDs across cameras and frames. This makes it the largest amodal segmentation dataset and the first to supply ground-truth amodal content labels rather than pseudo-labels derived from cut-and-paste schemes. A sympathetic reader would care because these labels make it possible to train and evaluate models for amodal content completion, multi-camera object tracking, and view-invariant object representation directly, without the noise of approximated targets.","feed_headline":"First ground-truth amodal content dataset spans 5.8M object instances","feed_subtitle":"Six cameras per scene with shared object IDs let models learn what occluded objects truly look like.","key_machinery":"The load-bearing object is the dataset itself, generated from a simulation engine: each of the 2,041 scenes is a two-second, 24-frame video captured by six cameras whose motion is static, linear, or an arc, and contains 2 to 40 generic household objects, including up to 20 objects thrown through the air to create heavy occlusion. For each object, the generator produces modal and amodal masks, amodal RGB content, and amodal depth, and assigns a consistent object ID across frames and cameras through scene-object descriptors. Those descriptors are what make the multi-camera consistency claim real: they let a model associate one object's appearance in all six views, which is the mechanism that could support view-invariant object representation and re-identification. The proposed metrics use the modal mask to subtract the visible region, so evaluation can focus on the occluded areas where completion is genuinely hard.","core_discovery":"On its own terms, the central discovery is the dataset: every object in every frame of every scene comes with modal masks, amodal masks, amodal RGB content (the 'x-ray' view through occluders), depth, and a scene-object descriptor that ties detections to a single identity across six cameras. With about 4.09 million occluded instances at an average occlusion rate of 45.2%, the dataset is large enough that occlusion-heavy training is feasible. The authors further propose multi-camera video amodal content completion as a new task and adapt PSNR, LPIPS, and SSIM, plus occluded-region variants, to measure how well a model reconstructs hidden content. Their argument is that this combination replaces the previous workaround of pseudo-label generation with direct supervision and adds camera perspective as a source of object context alongside time.","pith_inferences":["Editorial inference: the paper does not train any model, so a natural next experiment is to fine-tune an existing diffusion-based video amodal completion model on MOVi-MC-AC and compare it against the same model trained on pseudo-labels; if the ground-truth labels do not yield better occluded-region PSNR/LPIPS, the dataset's practical value would need stronger justification.","Editorial inference: the six cameras with shared IDs provide a free self-supervised signal: a model that must predict the same object's appearance from different viewpoints could learn view-invariant features that transfer to real-world re-identification, even though all training data is synthetic.","Editorial inference: because objects are rigid household items, claims about object permanence and amodal completion for articulated, deformable, or human objects are not yet tested; extending the simulator to such categories is a direct next step.","Editorial inference: the dataset is released without code to regenerate annotations, so an independent alignment check — re-rendering a sample of objects from their 3D models and comparing against the provided amodal RGB — would settle whether the ground truth is as clean as claimed."],"forward_implications":["Amodal content completion models can be trained directly on ground-truth occluded appearance rather than pseudo-labels from cut-and-paste schemes, making their outputs comparable for the first time.","Multi-camera object tracking and detection can be cast as a single task with one consistent ID per object across all six views, instead of per-camera tracking with later association.","Object retrieval and re-identification can be evaluated across cameras and across scenes with unseen object classes, because the training and test splits use disjoint object sets.","The occluded-region metric variants (e.g., PSNR_occ, LPIPS_occ, SSIM_occ) isolate reconstruction quality in the genuinely hidden parts of objects, preventing easy visible-region pixels from inflating scores.","The dataset's depth and amodal depth annotations open a path toward amodal 3D detection and amodal mask prediction beyond 2D image and video tasks."],"supporting_citations":[{"why":"The simulation engine that generates the multi-camera scenes, objects, and ground-truth annotations.","marker":"[39]"},{"why":"The diffusion-based video amodal completion method whose pseudo-label strategy the dataset replaces with ground truth.","marker":"[20]"},{"why":"A published video amodal completion model that the new task targets.","marker":"[21]"},{"why":"The source of the mIoU and mIoU_occ segmentation metrics adopted for amodal mask evaluation.","marker":"[15]"},{"why":"Another source for the mIoU_occ metric that focuses evaluation on occluded regions.","marker":"[41]"},{"why":"The LPIPS perceptual metric adapted here to score amodal content reconstruction.","marker":"[42]"},{"why":"The SSIM structural-similarity metric adapted here for reconstruction quality.","marker":"[43]"}],"fun_headline_variants":["Multi-camera amodal dataset with 5.8M instances reveals occluded content","First ground-truth amodal content dataset from multi-camera video","X-ray vision for machines: 5.8M occluded objects from six cameras","New benchmark: Multi-camera video teaches AI to see through occlusions","Largest amodal dataset yet: 5.8M instances with multi-camera views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ground-truth amodal content and the cross-camera object identities are produced correctly by the simulator and line up exactly with the RGB frames and masks; if the rendering is misaligned or an object's ID drifts between cameras, the dataset's central value collapses.","fun_headline_variants_meta":{"raw":{"variants":["Multi-camera amodal dataset with 5.8M instances reveals occluded content","First ground-truth amodal content dataset from multi-camera video","X-ray vision for machines: 5.8M occluded objects from six cameras","New benchmark: Multi-camera video teaches AI to see through occlusions","Largest amodal dataset yet: 5.8M instances with multi-camera views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1579,"prompt_tokens":1051,"completion_tokens":528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":423}},"tokens_in":667,"tokens_out":528,"duration_ms":6156,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:17:30.234827+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could take a random sample of occluded instances, independently re-render each object's unoccluded appearance from its 3D model, and compare pixel-by-pixel with the provided amodal content; any systematic misalignment, or any object whose ID is not present in all six cameras' annotation files, would falsify the dataset's central claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The simulation engine that generates the multi-camera scenes, objects, and ground-truth annotations."},{"cited_title":"Using diffusion priors for video amodal segmen- tation, 2024","cited_arxiv_id":null,"evidence_quote":"The diffusion-based video amodal completion method whose pseudo-label strategy the dataset replaces with ground truth."},{"cited_title":"TACO: Taming Diffusion for in-the-wild Video Amodal Completion","cited_arxiv_id":"2503.12049","evidence_quote":"A published video amodal completion model that the new task targets."},{"cited_title":"Rethinking amodal video segmentation from learning supervised signals with object-centric representa- tion, 2023","cited_arxiv_id":null,"evidence_quote":"The source of the mIoU and mIoU_occ segmentation metrics adopted for amodal mask evaluation."},{"cited_title":"Self-supervised amodal video object segmentation, 2022","cited_arxiv_id":null,"evidence_quote":"Another source for the mIoU_occ metric that focuses evaluation on occluded regions."},{"cited_title":"Efros, Eli Shechtman, and Oliver Wang","cited_arxiv_id":null,"evidence_quote":"The LPIPS perceptual metric adapted here to score amodal content reconstruction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The SSIM structural-similarity metric adapted here for reconstruction quality."}],"review_version":1}