{"id":"146b53a9-f4e6-48fb-a493-b67d96ed7d14","arxiv_id":"2412.15150","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Adding the HSV saturation channel to the RGB reconstruction target improves Slot Attention object discovery and disentanglement across several multi-object datasets.","lead":"This paper shows that object-centric models segment scenes better when the reconstruction target adds the saturation channel from HSV to the usual RGB channels, a nearly free change. The simple trick improves object discovery and disentanglement on five benchmark datasets, suggesting the output color space matters more than previously assumed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RGB-S gains are not isolated from a dimensionality confound — no control with an extra arbitrary channel is run, so the saturation-specific mechanism claimed in §3.3 is untested.","rationale":"After reading carefully, the paper's central empirical finding — RGB-S targets improve FG-ARI/mIoU on five datasets — is plausibly real and consistently reported; the lighting-robustness analysis in Tables D7 and D8 provides independent support that saturation carries information stable across illumination. However, the paper's central conceptual claim is that this improvement is due to saturation's specific properties (non-correlation, continuity, robustness to lighting). That claim is load-bearing because it is the stated reason for choosing RGB-S over other possible extensions and the basis for the general recommendation to use composite color spaces. The experimental design cannot discriminate this mechanism from a simple dimensionality effect, since no condition adds one extra channel with different content. The reader's weakest assumption identifies exactly this gap, and I agree. A single additional control — RGB plus a duplicate R channel — would settle it. If that control matches RGB-S, the paper's contribution reduces to 'predict more targets,' which is far less novel; if it does not, the saturation narrative is strengthened. This does not change the conditional verdict: the paper's empirical result is interesting, but the stated mechanism remains unproven until the control is added or the claims are softened.","tokens_in":23415,"tokens_out":5867,"duration_ms":55081,"concrete_test":"Run the same CNN architecture and hyperparameters on Clevrtex and MultiShapeNet-4 with an RGB-Rdup target (RGB plus a duplicate of the red channel, same 5-channel decoder, same 10 seeds). Also run RGB-const (fourth channel fixed to 0.5) as a second control. If RGB-Rdup or RGB-const attains FG-ARI/mIoU within the reported error band of RGB-S, the saturation-specific mechanism is falsified and the dimensionality confound is confirmed. Report seed-level FG-ARI for all conditions and a paired bootstrap test (n=10) comparing RGB-S against RGB-Rdup and against RGB, with the p-value and 95% confidence interval for the mean difference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 motivates RGB-S by saturation's non-correlation, continuity, and lighting robustness, and Table 2 interprets RGB-S's gains as validating that choice. But every alternative target space in Table 2 differs from RGB in both channel count and channel semantics: RGB-HSV and RGB-SV add two channels, HSV replaces all three, and RGB-R and RGB-GRAY in Table 1 are full replacements rather than additions. There is no condition that adds one extra dimension with content lacking saturation's purported properties, such as a duplicate R or G channel or a constant channel. The observed gains could therefore come entirely from increasing the decoder's output dimensionality, rather than from saturation's specific statistical properties. This attribution is load-bearing: the paper's conceptual contribution is 'color channel independence' as a principle, and the conclusion claims RGB-S 'significantly outperforms RGB in any experiment without negative side effects.' A control with an arbitrary correlated channel would determine whether the principle or the extra dimension drives the results. The paper's own note that the best RGB seed on MultiShapeNet-4 reaches 84.1 FG-ARI, matching the RGB-S mean of 82.1, further shows that part of the reported improvement is a reduction in degenerate seeds rather than a uniform representational gain; without seed-level paired comparisons, the 'markedly better' claim is underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the color-space representation used as the reconstruction target in unsupervised object-centric learning with Slot Attention. It argues that RGB targets are suboptimal because the RGB channels are mutually correlated and sensitive to lighting, and it proposes composite target spaces, most prominently RGB-S (RGB plus the saturation channel from HSV), which requires only changing the decoder's output dimensionality. The authors evaluate RGB-S, RGB-SV, RGB-HSV, HSV, and RGB baselines on five datasets (CLEVR, MultiShapeNet-4, MultiShapeNet-24, ClevrTex, MOVi-C) and report consistent mean improvements in FG-ARI and mIoU, together with DCI-based disentanglement metrics and an additional lighting-robustness experiment in Appendix D.4. The central claim is that RGB-S 'significantly' outperforms RGB and yields better object discovery and slot representations.","tokens_in":23624,"tokens_out":4844,"duration_ms":36336,"significance":"If the empirical claim holds, this is a practically useful and remarkably cheap improvement: it is architecture-agnostic, adds essentially no computational cost, and improves mean object-discovery metrics across several standard benchmarks. The paper has real strengths: experiments are run over 10 seeds for CNN models and 6 for ResNet models, evaluation is on held-out test sets, and the appendix includes out-of-distribution, lighting-robustness, and slot-ablation analyses. However, the conceptual contribution depends on attributing the gains to saturation's specific statistical properties, and this attribution is currently confounded with output dimensionality. The statistical strength of the headline claims is also overstated relative to the reported variances. These issues are fixable with additional analyses, but they are load-bearing for the paper's central interpretation.","major_comments":[{"comment":"The causal attribution of the gains to saturation's statistical properties (non-correlation, continuity, lighting robustness) is not isolated from a dimensionality confound. Every alternative target considered changes the channel count as well as the channel semantics: RGB-S adds one channel, RGB-SV and RGB-HSV add two, HSV replaces all three, and RGB-R and RGB-GRAY are full replacements. There is no condition that adds one extra dimension whose content lacks the claimed properties, such as a duplicate R channel, a constant channel, or a fixed random channel. Without such a control, the observed improvements could be explained by the increased decoder output dimensionality alone, and the paper's central conceptual claim about 'color channel independence' is unsupported. The authors should add an ablative control with an arbitrary extra channel and report whether the saturation-specific benefit remains.","section":"Section 3.3; Table 2"},{"comment":"The word 'significantly' is used repeatedly (Abstract, Sections 3.3, 5.1.2, 5.2, Conclusion), but no statistical significance test, confidence interval, or paired comparison is reported. Several headline differences are within one standard deviation of the baseline, e.g., CLEVR RGB-S vs RGB FG-ARI 94.9 ± 1.2 vs 94.1 ± 1.1 and MOVi-C RGB-S vs RGB FG-ARI 45.7 ± 2.4 vs 42.9 ± 3.3. Because the same seeds are used across target spaces, paired tests or seed-level effect sizes should be reported to substantiate the 'significant' wording; otherwise the conclusion should be softened to 'improvement in average metrics' rather than statistical significance.","section":"Section 5.1; Table 2"},{"comment":"The paper itself states that part of the RGB baseline's weakness comes from degenerate seeds (2 of 10 on MultiShapeNet-4, 5 of 10 on MultiShapeNet-24) and that the best RGB seed (84.1 FG-ARI on MultiShapeNet-4) matches the RGB-S mean (82.1). This suggests the improvement may be largely variance reduction rather than a uniform representational gain. The authors should report per-seed paired differences for every dataset, state how many runs degenerated per target space, and show that the advantage persists when degenerate runs are excluded. Without this, the 'markedly better' claim in the Abstract is underdetermined.","section":"Section 5.1.2"}],"minor_comments":[{"comment":"Raw citation commands such as 'citeBiza2023', 'citeMajellano2024', and 'citesingh2023' appear in the text and should be converted to proper references.","section":"Section 4"},{"comment":"The caption 'RGB2X' is unclear; the text describes RGB-R and RGB-GRAY as full replacements rather than two-channel additions, so the table header should be reworded.","section":"Table 1"},{"comment":"The text uses 'd=0' while the corresponding tables use 'L=0' and 'L=20'; the notation should be unified.","section":"Appendix D.8"},{"comment":"The caption refers to '(D7)' when comparing with HSV feature maps; it should refer to Figure D8.","section":"Figure D7"},{"comment":"The reference list contains duplicate entries for Greff et al. (2020) and several incomplete reference strings (e.g., Eastwood and Williams without a year or venue); these should be cleaned.","section":"References"},{"comment":"The RGB-MSE caveat is acknowledged, but because the compared models reconstruct different target spaces, the RGB-MSE values are not directly comparable across columns; reporting reconstruction loss in the respective target space would make the comparison cleaner.","section":"Table 2; Section 5"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is reasonably solid, and the RGB-S idea is simple enough to be a useful practical contribution if the attribution is supported. The missing arbitrary-channel control and the lack of seed-level significance testing are the two points most likely to be raised by other reviewers; I recommend making both mandatory for revision. I do not see a novelty or scope problem with this manuscript for a vision/representation-learning venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The empirical core of this paper is real: switching the reconstruction target from RGB to RGB-S improves FG-ARI and mIoU on all five datasets, and the ClevrTex gains are large. That is a new and practically useful result for object-centric learning. The paper is also honest about a key wrinkle: on MultiShapeNet-4 the best RGB seed reaches 84.1 FG-ARI, matching the RGB-S mean, so part of the reported improvement is about avoiding degenerate seeds rather than a uniform representational gain.\n\nWhat it does well: ten seeds for CNN and six for ResNet, consistent improvements, a lighting-robustness test on held-out scenes, DCI disentanglement analysis, out-of-distribution evaluation, and a comparison against DINO-based self-supervision. The limitations section is candid, and the paper correctly notes that the design space of composite channels is barely explored. This is a well-run empirical study, even without released code.\n\nThe main soft spot is exactly the one the stress test flags: the mechanism is not isolated. Every target in the tables differs from RGB in both channel count and channel semantics, so the gains could in principle come from simply giving the decoder an extra output dimension. There is no control with RGB plus an arbitrary channel (e.g. duplicate R or a constant). The conceptual story in Section 3.3, that saturation helps because it is uncorrelated and lighting-robust, is plausible but untested by that design. The repeated use of \"significantly\" without statistical tests is a smaller issue, but real.\n\nNone of this is fatal. The observation that output color space affects unsupervised object discovery is novel and appears robust across datasets. The missing control is the key addition I would ask a referee to require, along with significance testing and code release.\n\nSend it to peer review. It deserves a serious referee.","headline":"RGB-S targets reliably improve slot-attention object discovery, but the paper never isolates whether the gain comes from saturation's specific properties or merely from adding an extra output channel.","tokens_in":24197,"tokens_out":3121,"would_cite":true,"duration_ms":27740,"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":"Adding a saturation channel to the reconstruction target of Slot Attention models improves unsupervised object discovery across five multi-object datasets.","keywords":["object-centric representation learning","slot attention","unsupervised object discovery","color space","saturation","composite color spaces","reconstruction loss","disentanglement"],"falsifier":"Train the same Slot Attention architecture with RGB plus a control fourth channel, such as a duplicated R channel, a constant channel, or random noise, and compare FG-ARI and mIoU on Clevrtex and MultiShapeNet; if any control matches or exceeds RGB-S, the saturation-specific explanation fails, while the paper's account predicts RGB-S wins and the controls behave like RGB.","tokens_in":23196,"feed_emoji":"🎨","tokens_out":7114,"duration_ms":59826,"temperature":0.7,"pith_summary":"The paper asks whether the color space in which an unsupervised object-centric model is asked to reconstruct a scene should always be RGB. It argues that RGB channels are highly correlated and dominated by lightness, so a reconstruction loss gives slots little signal for separating objects that differ in hue or saturation. The proposed fix is a composite target space, RGB-S, that appends the HSV saturation channel to RGB in the decoder output. Training Slot Attention models to predict this extra channel improves object segmentation and factor disentanglement on all five tested datasets, with no architectural change beyond the decoder head and no extra compute. If correct, this means a free choice of output color channels can partially substitute for supervised or self-supervised signals.","feed_headline":"RGB plus saturation beats RGB alone at object discovery","feed_subtitle":"Asking Slot Attention models to predict HSV's saturation channel alongside RGB improves segmentation on all five tested datasets, for free.","key_machinery":"The load-bearing object is the composite RGB-S target color space: the decoder is asked to predict RGB channels plus the saturation channel and the alpha mask, and the unsupervised reconstruction loss includes saturation as an additional target. Saturation is continuous, nearly uncorrelated with all three RGB channels, and robust to illumination changes, so it supplies a complementary signal that forces slots to represent perceptual properties that RGB's lightness-dominated channels underweight. The argument runs through the slot-binding mechanism: with the three RGB channels effectively acting as one correlated lightness dimension, the attention competition has too little to discriminate on; the fifth uncorrelated dimension gives it a foothold.","core_discovery":"The discovery is that the target representation, not just the input, controls what slot-based models learn to bind. The authors show that when the spatial-broadcast decoder predicts an extra channel, saturation from the HSV space, alongside RGB, object discovery (FG-ARI and mIoU) and the ability of slots to encode object properties improve consistently. Pure HSV targets work well on textured datasets but degenerate on simpler ones because the hue channel is discontinuous; RGB-S keeps RGB's stability and adds only the continuous, RGB-uncorrelated saturation signal. In their experiments, RGB-S and RGB-SV outperform RGB in every setting, with Clevrtex FG-ARI rising from 75.6 to 92.7 and Movi-C mIoU from 21.3 to 27.2, while precision for predicting object materials roughly doubles.","pith_inferences":["A control condition with an extra channel that is not semantically motivated, such as a duplicated or noise channel, would test whether the benefit comes from saturation's statistical properties or merely from increased target dimensionality; no such control appears in the paper.","If the RGB-S mechanism is what the paper thinks it is, it should combine with self-supervised objectives rather than compete with them: saturation targets provide complementary information, whereas self-supervised features on Clevrtex underperform composite spaces in the paper's comparison.","A testable extension is to synthesize scenes where saturation of objects is systematically decorrelated from their RGB lightness; the paper's account predicts object discovery should improve most exactly when saturation alone separates objects from the background.","Composite color targets may also improve other pixel-reconstruction learning tasks such as image colorization or autoencoding, since the same correlated-RGB-channels argument applies there; the paper leaves this untested."],"forward_implications":["Adding saturation to the reconstruction target improves foreground-adjusted Rand index and mIoU over RGB across all five datasets, with the largest gains on textured scenes such as Clevrtex.","Slots trained with RGB-S targets better encode underlying object properties: average precision for predicting Clevrtex materials roughly doubles compared with RGB, and DCI informativeness, disentanglement, and completeness all improve.","Slot representations trained with saturation targets are more stable under changing lighting conditions, as measured by smaller cosine and Euclidean distances for the same object under darkness shifts.","The benefit comes at nearly zero cost: only the decoder's output dimension and the loss target change, so the recipe applies to any slot-based or reconstruction-trained visual model.","On simple datasets, adding hue destabilizes training, so saturation rather than any extra HSV channel is the channel that carries the consistent improvement."],"supporting_citations":[{"why":"Introduces the Slot Attention module and the RGB-reconstruction training setup that all experiments in this paper modify.","marker":"Locatello et al. (2020)"},{"why":"Documents the failure of pure RGB reconstruction loss for real-world datasets such as Movi-C and supplies baseline numbers and a self-supervised comparison.","marker":"Seitzer et al. (2023)"},{"why":"Provides the invariant Slot Attention architecture used for MultiShapeNet, Clevrtex, and Movi-C, preventing trivial spatial binding.","marker":"Biza et al. (2023)"},{"why":"Supplies the Clevrtex dataset and its out-of-distribution test set, the textured benchmark where RGB-S shows its largest gains.","marker":"Karazija et al. (2021)"},{"why":"Supplies the Kubric-based Movi-C dataset, the photorealistic video-frame benchmark where RGB degenerates into areas and RGB-S improves mIoU.","marker":"Greff et al. (2022)"},{"why":"Supplies the CLEVR dataset, the simple proof-of-concept benchmark where RGB-S slightly improves segmentation without degenerating.","marker":"Johnson et al. (2016)"},{"why":"Supplies the MultiShapeNet dataset and its 4- and 24-object variants, where RGB seeds degenerate and RGB-S reduces degeneration.","marker":"Stelzner et al. (2021)"},{"why":"Provides the HSV color space definition and the perceptual grounding for why saturation is continuous and complementary to RGB.","marker":"Smith (1978)"}],"fun_headline_variants":["Adding saturation to RGB improves object discovery","RGB-S color space improves object detection without extra cost","Predicting saturation channel enhances object discovery","Extra color channel in targets improves object binding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the improvement comes from saturation's specific statistical properties (uncorrelated with RGB, robust to lighting, continuous), not simply from the model being asked to predict an extra channel; the paper gives no control condition with an arbitrary fourth channel.","fun_headline_variants_meta":{"raw":{"variants":["Adding saturation to RGB improves object discovery","RGB-S color space improves object detection without extra cost","Predicting saturation channel enhances object discovery","Extra color channel in targets improves object binding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1579,"prompt_tokens":895,"completion_tokens":684,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":511,"tokens_out":684,"duration_ms":6450,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:34:28.056285+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Slot Attention architecture with RGB plus a control fourth channel, such as a duplicated R channel, a constant channel, or random noise, and compare FG-ARI and mIoU on Clevrtex and MultiShapeNet; if any control matches or exceeds RGB-S, the saturation-specific explanation fails, while the paper's account predicts RGB-S wins and the controls behave like RGB.","supporting_citations":[{"cited_title":", Weissenborn , D","cited_arxiv_id":null,"evidence_quote":"Introduces the Slot Attention module and the RGB-reconstruction training setup that all experiments in this paper modify."},{"cited_title":", Van Steenkiste , S","cited_arxiv_id":null,"evidence_quote":"Provides the invariant Slot Attention architecture used for MultiShapeNet, Clevrtex, and Movi-C, preventing trivial spatial binding."},{"cited_title":": Color gamut transform pairs","cited_arxiv_id":null,"evidence_quote":"Provides the HSV color space definition and the perceptual grounding for why saturation is continuous and complementary to RGB."}],"review_version":1}