{"id":"865e01d9-1c99-4e5b-912d-3ad1b23074fc","arxiv_id":"2412.12331","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Using CroCo's geometric features as both the encoder and reconstruction target improves object discovery in slot-attention video models, with a one-pass attentional decoder.","lead":"This paper proposes a video representation model that splits scenes into object slots using pre-trained geometric features from CroCo, and uses a new lightweight slot decoder to read out all objects in one pass. It reports better object segmentation than prior slot-based methods on synthetic MOVi videos while using much less memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central comparison in Table II confounds prediction target with decoder and pipeline; without a matched-target ablation, the geometric-prior claim is not established.","rationale":"The reader's weakest-assumption analysis matches my read of the paper: Table II is not a clean test of the geometric-prior hypothesis because target, decoder, and temporal pipeline vary across methods. The paper does provide useful independent support for the attentional slot decoder itself: Table I reports ablations (removing positional embedding or global scene embedding) and a memory reduction from 24 GB to 6 GB, which are concrete architectural contributions. However, those ablations do not isolate the prediction target. The paper also gives qualitative evidence (Figure 1) that CroCo attention is less object-like than DINO's, which makes the causal story more surprising rather than less. Since the central claim is explicitly stated as a superiority of geometric over semantic representations (Section IV-B), the burden is on a controlled comparison. The proposed test is one specific experiment that would settle whether the concern lands. I agree with the reader's CONDITIONAL verdict: the method is promising and the decoder contribution is credible, but the geometric-prior advantage is not yet established.","tokens_in":8777,"tokens_out":3357,"duration_ms":33533,"concrete_test":"Re-run the Section IV-B comparison with a single controlled pipeline: hold fixed the proposed attentional slot decoder, Slot Attention module, recurrent slot initialisation, bounding-box conditioning, optimizer, learning rate, and training budget, and vary only the prediction target between CroCo features and DINO features (plus, for completeness, MAE features). Report ARI-FG on MOVi-C and MOVi-E over at least three seeds as mean +/- std. If the CroCo-target run does not beat the DINO-target run by more than the seed noise, the geometric-prior claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that the geometric CroCo target drives the +10% ARI-FG gain over DINOSAUR in Table II—is not identifiable from the reported comparisons. The three methods differ jointly in prediction target, decoder architecture, and temporal-slot pipeline: DINOSAUR is an image model adapted with 'the same training pipeline', while the proposed method uses the attentional slot decoder from Section III-B plus recurrent slot initialisation inherited from SA Vi. No ablation holds the decoder and pipeline fixed and varies only the prediction target (CroCo vs DINO). The observed gain could therefore be explained by the decoder's cross-attention mechanism, the residual position injection, or the recurrent slot carry-over, rather than by geometric information. This concern is sharpened by the paper's own observation in Section III-A that CroCo attention maps are diffuse and show no object grouping: without a controlled target swap, the mechanism attributed to 'geometric representation' remains speculative. The absence of error bars or multiple-seed results further weakens the comparison, but the core issue is the confound between target choice and architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an object-centric video representation learning framework that uses a pre-trained CroCo encoder's geometric feature map as the reconstruction target for a Slot Attention module. The main architectural contribution is an Attentional Slot Decoder that uses a 2D Fourier-positional query grid, a global scene embedding, and cross-attention over slot latents, with the attention scores directly serving as segmentation masks. On MOVi-C and MOVi-E, the method reports ARI-FG values of 0.788 and 0.766, respectively, exceeding SA Vi and DINOSAUR and approaching the depth-supervised SA Vi++ (0.843/0.823), while requiring about 6 GB of training memory versus 24 GB for SA Vi. The paper also includes decoder ablations, qualitative results, and an analysis of attention maps from several pre-trained models.","tokens_in":8969,"tokens_out":5235,"duration_ms":51162,"significance":"If the results are robust, the paper would support a useful and somewhat counterintuitive claim: a geometric self-supervised representation (CroCo) can serve as a better reconstruction target for slot-attention object discovery than a semantic representation (DINO), and the proposed decoder gives a major memory saving. The paper has several concrete strengths: the decoder ablations in Table I (positional embedding, global scene embedding) isolate specific design choices, the evaluation uses external ground-truth masks and ARI-FG, so there is no training/evaluation circularity, and the paper honestly documents that CroCo attention maps do not show object grouping. However, the central causal claim about the geometric prior is currently supported only by a cross-method comparison that changes several variables at once, and the main quantitative comparisons are single-run point estimates. The potential value is real, but the evidence needs targeted additions before the claim is established.","major_comments":[{"comment":"The central claim that the geometric CroCo target, rather than the decoder or training pipeline, causes the +10% ARI-FG gain over DINOSAUR is not identifiable from the reported experiments. The rows of Table II differ jointly in prediction target, decoder architecture, and temporal slot-initialisation: DINOSAUR is adapted from its image version with its own decoder, while the proposed method uses the Attentional Slot Decoder from Section III-B together with the recurrent slot initialisation inherited from SA Vi. No ablation holds the decoder and pipeline fixed and varies only the prediction target. A minimal test would train the proposed pipeline with DINO features as the target and compare it to the CroCo-target version (or, symmetrically, train DINOSAUR's decoder with CroCo features). Without such an ablation, the improvement in Table II could be due to the decoder's cross-attention, the residual positional injection, or the recurrent slot carry-over, rather than to geometric information. This is load-bearing because the paper's title and Section IV-B explicitly attribute the improvement to geometric versus semantic representations.","section":"Section IV-B, Table II"},{"comment":"All reported values in Table II, including the headline margins (0.788 vs 0.686 on MOVi-C and 0.766 vs 0.651 on MOVi-E), are point estimates with no error bars, no number of seeds, and no significance test. Slot Attention training is known to be sensitive to seed variation, so the observed margins could shrink or change in relative ordering. The paper should report means and standard deviations over at least three seeds, or provide a clear statement of variance in the appendix. This is not merely a presentation issue because the claims 'we surpass' and '+10% improvement' are stated in absolute terms from single runs.","section":"Section IV-B, Table II"},{"comment":"The paper's own analysis in Section III-A(c) shows that CroCo attention maps are diffuse and do not exhibit object-grouping behavior, unlike DINO and MSN. This observation is not a direct argument against the geometric-prior hypothesis, but it also provides no supporting evidence for the mechanism by which geometric features improve object discovery. The paper says the improvement is 'attributed to the distinctive characteristics of geometric and semantic representations,' yet no experiment measures or ablates the geometric content of the target features. The matched-target ablation proposed in the first comment would be the minimal way to support this mechanism; without it, the mechanism remains speculative.","section":"Section III-A(c), Section IV-B"}],"minor_comments":[{"comment":"The row labels '- Pos.' and '- CLS' are not explained in the table caption. The text describes removing positional information and removing the global scene representation, but the table abbreviates the latter as 'CLS'; please define whether 'CLS' refers to the global embedding taken from the CLS token and confirm the correspondence with the text.","section":"Table I"},{"comment":"The SA Vi++ values (0.843 and 0.823) are listed without a source citation in the table. Please state whether these numbers are taken directly from [17] or re-evaluated under the same protocol, and cite the specific source or appendix.","section":"Table II"},{"comment":"The abstract calls the method 'weakly-supervised,' but the paper also describes the task as 'unsupervised discovery' and uses bounding boxes only for conditional slot initialisation. Please clarify the terminology and reconcile it with the conditional object-centric learning setting inherited from SA Vi.","section":"Abstract and Section IV"},{"comment":"The loss is written as L = ||Y - V||_2^2 without stating whether the norm is averaged over spatial positions and feature channels. Please clarify the exact MSE normalization used in training.","section":"Section III-C"},{"comment":"There are two typos in the conclusion: 'bootstraping' should be 'bootstrapping,' and 'compatability' should be 'compatibility.'","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is a solid contribution with a clear efficiency advantage and a plausible hypothesis, but the central claim about the geometric prior currently rests on a confounded comparison and single-run numbers. If the authors add a matched-target ablation (same decoder/pipeline, CroCo vs DINO targets) and multi-seed statistics, the paper could become suitable for publication. I do not see any concern about novelty disclosure or citation ethics; the related work covers the relevant baselines. The paper would also benefit from releasing code to support the reproducibility of the reported ARI-FG values."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent empirical paper giving the object-centric video community two useful things—a one-pass attentional slot decoder that cuts training memory roughly 4x (24GB to 6GB), and evidence that a reconstruction target with geometric signal (CroCo features) beats a semantic one (DINO/DINOSAUR) on MOVi-C/E object discovery. The decoder is a sensible set-based cross-attention variant with positional queries plus a broadcast global embedding; the memory reduction is real, and the ablations show the positional and global-embedding choices matter.\n\nThe result is more interesting because the paper shows CroCo's attention maps are diffuse and non-object-like, yet slot attention on CroCo features yields higher ARI-FG than DINO features. That counterintuitive finding is worth taking seriously.\n\nThe soft spot is the central comparison. Table II varies prediction target, decoder architecture, and temporal pipeline jointly. DINOSAUR is adapted from its image version with its own decoder, while the proposed method uses the attentional decoder plus recurrent slot initialization inherited from SA Vi. No ablation holds the decoder fixed and swaps only CroCo vs DINO as target. So the observed +10% ARI-FG over DINOSAUR could come from the decoder's cross-attention mechanism or the recurrent slot carry-over, not from geometric information. The paper does include useful ablations of its own decoder, but not this key target swap. There are also no error bars or multiple seeds; the margins are large, but seed variation is unquantified.\n\nMinor points: SA Vi with RGB pixels is a weak baseline, so beating it is not surprising; the meaningful comparisons are DINOSAUR and SA Vi++. The paper is honest about failure cases and error accumulation. The citation pattern is appropriate, with no self-citation inflation.\n\nBottom line: the empirical result is probably in the right direction, but the headline claim is conditional. It deserves a serious referee: send it to review, and require the matched-decoder target ablation, multi-seed results, and ideally code. If the target swap confirms the geometric-prior advantage, it becomes a solid contribution.\n\nFor whom: researchers working on slot attention, object-centric video, and self-supervised feature targets. Recommendation: engage, with the ablation as the gate.","headline":"A useful, counterintuitive empirical result on CroCo features for object discovery, but the central geometric-prior claim is confounded with decoder and pipeline changes—needs a matched-target ablation and error bars before accepting.","tokens_in":9485,"tokens_out":2647,"would_cite":true,"duration_ms":23790,"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":"This paper claims that a pre-trained geometric vision model, used as the reconstruction target for slot attention, improves object-centric segmentation in complex synthetic video, outperforming a semantic target and approaching…","keywords":["object-centric representation learning","slot attention","geometric prior","self-supervised pre-training","video object segmentation","attentional slot decoder","weakly supervised learning"],"falsifier":"Train the paper's exact pipeline but replace the CroCo feature target with DINO features (or vice versa, train DINOSAUR's pipeline with CroCo features) while keeping every other component identical. If ARI-FG does not move with the target, the geometric-prior claim is falsified; if it does, the claim is supported.","tokens_in":8576,"feed_emoji":"🧩","tokens_out":5154,"duration_ms":41157,"temperature":0.7,"pith_summary":"The paper argues that the target a slot-attention video model is trained to reconstruct determines how well it discovers objects, and that geometric features are a better target than semantic features. Training an object-centric video model to predict CroCo's pre-trained geometric representations, rather than DINO's semantic representations, raises foreground segmentation ARI-FG from 0.686 to 0.788 on MOVi-C and from 0.651 to 0.766 on MOVi-E, close to the depth-supervised SA Vi++ (0.843 and 0.823). This is achieved without explicit depth input, using a newly designed attentional slot decoder that performs a single decoding pass and reduces training memory from about 24 GB to about 6 GB. The practical stake is that object-centric learning, which has struggled to scale to complex scenes, might be bootstrapped from readily available geometric self-supervised models rather than requiring dense supervision.","feed_headline":"Geometric features beat semantic ones for object discovery","feed_subtitle":"CroCo features as target nearly match depth supervision with a quarter of the memory.","key_machinery":"The load-bearing component is the Attentional Slot Decoder. Instead of a spatial broadcast decoder that reconstructs each slot independently, this decoder initialises a spatial grid of Fourier positional embeddings, adds a pooled global scene embedding, and runs one cross-attention operation in which the slot latents serve as keys and values. The softmax-normalised attention scores over the slots are directly used as alpha masks for segmentation, and a residual connection from positional embeddings lets the decoder place objects at the correct locations. Around this, the method uses Slot Attention as the slot encoder, CroCo's pre-trained ViT as both visual backbone and reconstruction target, and a mean-squared-error loss in the feature space. The efficiency gain comes from replacing the per-slot decoding passes of mixture-of-components decoders with a single set-based cross-attention pass.","core_discovery":"The central claim is that the inherent geometric information in a pre-trained cross-view completion model (CroCo) provides a better training signal for object-centric discovery than the semantic information in DINO. Counterintuitively, even though DINO's attention maps visibly focus on foreground objects while CroCo's attention is diffuse, models trained to reconstruct CroCo features segment objects more accurately. The paper attributes this to geometry encoding 3D layout, occlusion, and motion cues that semantics lack. On MOVi-C and MOVi-E, the proposed method surpasses DINOSAUR (which uses DINO features) by roughly 10 absolute ARI-FG points and approaches SA Vi++, which uses explicit depth supervision. The authors also claim their attentional slot decoder reaches this performance with a single forward decoding pass and roughly a quarter of the training memory.","pith_inferences":["A direct test that would isolate the geometric prior's contribution is to train the same attentional slot decoder with DINO and with CroCo targets under identical slot initialisation and training schedule; the paper's comparison does not fully control for decoder architecture, so part of the +10% may come from the decoder rather than the geometric prior.","If geometry is the key, then other geometric or 3D-aware pre-training objectives—for example depth-prediction or multi-view consistency—should produce comparable or better object discovery than CroCo, offering a cheap route to scale object-centric learning to real-world video.","The inflated object masks the paper observes (shadows included) hint that geometric features encode object extent rather than tight boundaries; fine-tuning the target with contour or edge information might sharpen segmentation without losing the geometric advantage."],"forward_implications":["If correct, object-centric video models can be trained to state-of-the-art segmentation without explicit depth, using only frozen pre-trained geometric features as the target.","The attentional slot decoder reduces training memory roughly fourfold relative to SA Vi's spatial broadcast decoder, which lowers the barrier to scaling slot attention to longer or higher-resolution videos.","The choice of reconstruction target—geometric versus semantic—can matter more than the objectness apparent in the pre-trained model's attention maps.","The method is modality-flexible: the same decoder works with RGB, optical flow, and pre-trained feature targets, suggesting a single architecture can suit various reconstruction signals."],"supporting_citations":[{"why":"Supplies the CroCo pre-trained geometric features that serve as the reconstruction target and backbone, the central element whose benefit the paper demonstrates.","marker":"[14]"},{"why":"DINOSAUR is the semantic-target baseline that the paper improves upon; it established the practice of learning object-centric representations from pre-trained feature targets.","marker":"[18]"},{"why":"SA Vi provides the conditional object-centric video pipeline (bounding-box slot initialisation and recurrent slot carry-over) that the paper adapts, and its spatial broadcast decoder is the memory baseline.","marker":"[6]"},{"why":"SA Vi++ is the depth-supervised method whose ARI-FG scores mark the upper end the paper approaches without explicit depth.","marker":"[17]"},{"why":"Slot Attention is the module used to obtain object latents from the visual feature map, the core of the encoder side of the architecture.","marker":"[5]"},{"why":"DINO is the pre-trained semantic model whose features form the DINOSAUR target and whose attention maps serve as the semantic-contrast in the paper's analysis.","marker":"[12]"}],"fun_headline_variants":["CroCo beats DINO for object discovery without depth","Geometric prior rivals depth supervision in object learning","Slot decoder with CroCo features: cheaper, better segmentation","Object-centric learning: geometry trumps semantics","Pre-trained geometric prior boosts object discovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results hinge on the premise that the segmentation gains come from the geometric prediction target rather than from the new attentional decoder or the recurrent slot initialisation, since the comparison against DINOSAUR changes both the target and the architecture.","fun_headline_variants_meta":{"raw":{"variants":["CroCo beats DINO for object discovery without depth","Geometric prior rivals depth supervision in object learning","Slot decoder with CroCo features: cheaper, better segmentation","Object-centric learning: geometry trumps semantics","Pre-trained geometric prior boosts object discovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1145,"prompt_tokens":791,"completion_tokens":354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":407,"tokens_out":354,"duration_ms":3731,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:11:47.972798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the paper's exact pipeline but replace the CroCo feature target with DINO features (or vice versa, train DINOSAUR's pipeline with CroCo features) while keeping every other component identical. If ARI-FG does not move with the target, the geometric-prior claim is falsified; if it does, the claim is supported.","supporting_citations":[{"cited_title":"CroCo: Self-supervised pre- training for 3D vision tasks by cross-view completion,","cited_arxiv_id":null,"evidence_quote":"Supplies the CroCo pre-trained geometric features that serve as the reconstruction target and backbone, the central element whose benefit the paper demonstrates."},{"cited_title":"Bridging the Gap to Real-World Object-Centric Learning.,","cited_arxiv_id":null,"evidence_quote":"DINOSAUR is the semantic-target baseline that the paper improves upon; it established the practice of learning object-centric representations from pre-trained feature targets."},{"cited_title":"SA Vi++: Towards end-to-end object-centric learning from real-world videos,","cited_arxiv_id":null,"evidence_quote":"SA Vi++ is the depth-supervised method whose ARI-FG scores mark the upper end the paper approaches without explicit depth."},{"cited_title":"Object-Centric Learning with Slot Attention.,","cited_arxiv_id":null,"evidence_quote":"Slot Attention is the module used to obtain object latents from the visual feature map, the core of the encoder side of the architecture."},{"cited_title":"Emerging properties in self-supervised vision transformers,","cited_arxiv_id":null,"evidence_quote":"DINO is the pre-trained semantic model whose features form the DINOSAUR target and whose attention maps serve as the semantic-contrast in the paper's analysis."}],"review_version":1}