{"id":"67333f5f-c685-4449-b725-13d475b05526","arxiv_id":"2509.02545","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"MR-DINOSAUR refines the pre-trained DINOSAUR model with pseudo-labels derived from unsupervised optical flow on quasi-static video frames to achieve state-of-the-art unsupervised multi-object discovery on TRI-PD and KITTI.","lead":"A computer vision method, MR-DINOSAUR, discovers object instances in driving videos without any human labels by using motion from video frames where the camera is still. It beats previous approaches on two benchmarks and is one of the first fully unsupervised methods to do so.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hand-set pseudo-label thresholds (Sec. 3.1) are tuned per dataset with no stated validation procedure, leaving the 'fully unsupervised' SOTA claim fragile to threshold selection.","rationale":"The reader's weakest_assumption is exactly the quasi-static frame retrieval and the per-dataset thresholds; my analysis agrees that this is the most load-bearing concern. The paper provides strong empirical evidence (well-ablated components, direct pseudo-label comparison, and a validated quasi-static retrieval method) that the method works, but the central claim of fully unsupervised SOTA is contingent on threshold selection. The missing sensitivity analysis and validation procedure is a real soft spot, but it does not overturn the results as presented; it makes the claim conditional, matching the reader's CONDITIONAL verdict. I recommend no verdict change, while emphasizing the need for threshold robustness checks.","tokens_in":19422,"tokens_out":10562,"duration_ms":101001,"concrete_test":"On TRI-PD, fix all other hyperparameters and sweep τfg over {1.5, 2.0, 2.5, 3.0, 3.5} and τstatic over {0.3, 0.5, 0.7} on a held-out validation split (e.g., 5% of training videos), reporting final F1_50 after stage 1+2. Additionally, repeat the KITTI experiment with τstatic = 1.2 and 2.2, and τfg = 2.0 and 3.0. If F1_50 varies by more than 5 points across the plateau around the chosen values, the headline result is threshold-sensitive and the 'fully unsupervised' claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MR-DINOSAUR is a fully unsupervised method that outperforms prior SOTA is load-bearing, and the weakest link is the pseudo-label generation pipeline (Sec. 3.1). The pipeline depends on three hand-set thresholds: τstatic (quasi-static frame retrieval, per-dataset values 0.5 for TRI-PD and 1.7 for KITTI), τfg (foreground flow magnitude, 2.5 for both), and τ∇ (flow-gradient threshold, 20). The paper reports no sensitivity analysis and no procedure for choosing these thresholds without labels. If the thresholds were selected using the evaluation benchmark's ground truth (or even via qualitative tuning that implicitly relies on scene priors), the 'fully unsupervised' claim is weakened: the method effectively incorporates benchmark-specific information. Moreover, without a robustness sweep, we cannot rule out that the reported F1_50 gains (48.1 vs 41.5 on TRI-PD; 35.0 vs 23.2 on KITTI) sit on a sharp peak—small threshold changes could collapse performance. The supplementary validation of quasi-static retrieval (Tab. 7, 99.4% accuracy on KITTI) addresses only τstatic, not τfg or τ∇. The lower fg-ARI on MOVI-E (80.1 vs DIOD's 82.2) confirms that the static-camera assumption limits applicability, but the central benchmark claims do not directly test this fragility.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MR-DINOSAUR, an unsupervised multi-object discovery method that extends DINOSAUR by refining its slot representations with pseudo-labels derived from motion segmentation of quasi-static video frames. Pseudo-labels are generated by thresholding unsupervised optical flow (SMURF) on frames with minimal camera motion, extracting connected components, and splitting them via flow-gradient thresholding and HDBSCAN. Stage 1 fine-tunes the DINOSAUR slot-attention module with a weighted binary cross-entropy loss between matched slot alpha masks and pseudo-instance masks; stage 2 trains a slot deactivation module with a foreground/background loss and a similarity-based drop-loss to handle static objects. The method is evaluated on TRI-PD, KITTI, and MOVI-E, achieving F1_50 of 48.1 on TRI-PD and 35.0 on KITTI, improving over the previous supervised-pseudo-label method DIOD (41.5 and 23.2).","tokens_in":19716,"tokens_out":10092,"duration_ms":91193,"significance":"If the results hold, this is a valuable contribution to unsupervised multi-object discovery. The paper demonstrates that high-quality pseudo-labels can be obtained without any supervision by exploiting quasi-static frames, replacing the supervised TSAM pseudo-labels used by prior work. The two-stage training scheme and the slot-deactivation module are simple and effective, and the drop-loss is a sensible mechanism for recovering static objects. The paper includes useful ablations (Table 3), a pseudo-label quality analysis (Table 4), and an equal-footing comparison with DIOD retrained on KITTI (Table 2, DIOD*‡). The code release and detailed hyperparameters in the supplement support reproducibility.","major_comments":[{"comment":"The pseudo-label pipeline depends on three hand-set thresholds (τstatic, τfg, τ∇). The supplementary validates only τstatic, and only on KITTI (Tab. 7); no sensitivity analysis is provided for τfg or τ∇, and no validation is reported for τstatic on TRI-PD. Moreover, on MOVI-E (supp. D.1), which has constant camera motion, the method retrieves 84,831 of 87,741 frames as quasi-static, indicating that the static-background assumption is effectively not enforced there. Because the 'fully unsupervised' claim is load-bearing, the authors should state how each threshold was chosen without using benchmark labels and provide a robustness sweep, such as F1_50 as a function of τfg and τ∇ over a reasonable range, to show that the reported gains over DIOD are not peaked.","section":"§3.1, Tabs. 7–8"},{"comment":"The drop-loss is the only mechanism for discovering objects that are static in the current frame, but its effect is shown only globally (F1_50 46.3 → 48.1 in Tab. 3b). The claim that the method 'discovers objects beyond the dynamic ones' is not directly evaluated; the cosine-similarity condition (max c_u ≤ τdrop) may recover same-category static objects but could miss static objects of categories never observed moving. The authors should provide evidence for static-object discovery, for example by computing F1_50 separately for moving and static instances in the KITTI ground truth, and analyze the fraction of discovered objects that are static.","section":"§3.2, Eq. (4), Tab. 3b"}],"minor_comments":[{"comment":"All main results are reported without error bars or multiple seeds; please add at least a note on variance or run additional seeds for the key comparisons.","section":"§4.1, Tabs. 1–2"},{"comment":"The high AP50 (59.7) but low AR50 (24.7) on KITTI indicates a strong precision-recall imbalance; please discuss whether this is due to the conservative slot deactivation, and how it affects the interpretation of F1_50.","section":"§4.1, Tab. 2"},{"comment":"The supplementary lists the GPU as 'NVIDIA RTX 6000 Ada' while the main text says 'NVIDIA A6000 Ada'; please use consistent hardware names.","section":"Supplementary D.2"},{"comment":"The F1_50 gains are shown as '+6.6%' and '+11.8%', but the absolute values on the bars would help readers; please consider labeling them.","section":"Fig. 1"},{"comment":"The quasi-static frame retrieval uses corner patches covering 15% of image height and width; please justify this choice or note that it is dataset-agnostic.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the empirical work is solid for a workshop submission. The main risk is the hand-set thresholds: if they are not robust, the 'fully unsupervised' state-of-the-art claim weakens. The authors should be asked to provide a sensitivity analysis and a clearer statement of how thresholds were chosen. I would not reject on the basis of the unsupervised phrasing alone, since no labels are used in training."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: trust the central result, ask for a threshold sweep. This is a solid, well-ablated paper that mostly delivers on its claim. The authors take DINOSAUR, a pretrained slot-attention model, and refine it with pseudo-labels derived from unsupervised optical flow on quasi-static frames. Then they train a small slot deactivation MLP to suppress background slots. Simple as it sounds, it raises F1_50 from 41.5 to 48.1 on TRI-PD and from 23.2 to 35.0 on KITTI over DIOD* (the previous SOTA with DINOv2). The ablation table is convincing: stage 1 improves slot binding on its own, stage 2 produces the large jump, and the drop-loss gives a small consistent bump. The equal-footing comparison, where DIOD is retrained on KITTI without TRI-PD pretraining, is the right control and shows a 20.9 point F1 gap. They also compare their pseudo-labels against TSAM labels and win clearly on the shared subset while using no supervision. Code is public, which matters.\n\nThe soft spots are real but not fatal. The main one, which the stress-test note gets right, is that the pseudo-label pipeline depends on three hand-set thresholds (tau_static, tau_fg, tau_grad) and the paper reports no sensitivity analysis. If those values were tuned on the benchmark test set—or even selected by qualitative inspection that implicitly knows what objects look like—the 'fully unsupervised' SOTA claim gets weaker. I don't think that's the case: the thresholds are simple, the margins are large on KITTI, and the method degrades gracefully on MOVI-E. But the authors should show a small sweep over each threshold and state how the values were chosen. The absence of error bars is a minor weakness, standard for this area but worth fixing. Also, the supplement's MOVI-E section is internally confusing: it says constant camera motion violates the static-frame assumption, yet reports retrieving 84,831 'quasi-static' frames out of 87,741. That needs clarification.\n\nOverall, the central claim holds. The method is genuinely unsupervised in the sense that prior work is not: DIOD and BMOD rely on TSAM, which is supervised. The limitation to moving objects and static cameras is stated clearly. This is a paper for people working on object-centric learning, unsupervised instance segmentation, or driving-scene perception. It deserves a serious referee. I'd recommend acceptance conditional on a threshold sensitivity analysis and error bars; the core contribution is solid.","headline":"A well-ablated, genuinely unsupervised extension of DINOSAUR that beats prior SOTA on TRI-PD and KITTI; trust the central result, but ask for a threshold sensitivity sweep.","tokens_in":20243,"tokens_out":4484,"would_cite":true,"duration_ms":40426,"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":"Motion-only pseudo labels let unsupervised slot attention beat supervised baselines on driving scenes.","keywords":["unsupervised multi-object discovery","object-centric learning","slot attention","motion segmentation","optical flow","pseudo-label generation","slot deactivation","quasi-static frames"],"falsifier":"Run MR-DINOSAUR on a video dataset where every frame has camera motion, once with the still-frame filter and once with it turned off; if the two versions perform equally, the static-camera assumption is not load-bearing, whereas the paper's own results on the constantly-moving MOVI-E dataset suggest the filtered version should win.","tokens_in":19228,"feed_emoji":"🚗","tokens_out":9652,"duration_ms":83859,"temperature":0.7,"pith_summary":"MR-DINOSAUR sets out to show that multi-object discovery—detecting and segmenting every object instance in a scene—can be done with no human labels at all, even though prior methods quietly rely on supervised motion-segmentation masks for their pseudo labels. The paper's bet is that a pre-trained slot-attention model can be transferred to this task using only motion cues: retrieve video frames with a static camera, threshold unsupervised optical flow to find moving objects, and cluster the flow into instance masks. Those masks refine the slot representations, and a trained slot-deactivation module decides which slots are objects and which are background. If correct, a minimal, fully unsupervised pipeline outperforms the previous state of the art, as it does on TRI-PD and KITTI ($F_1$ at 50% IoU: 48.1 vs 41.5, and 35.0 vs 23.2). The cost is that the pseudo labels presuppose a quasi-static camera and moving objects, so datasets with persistent camera motion blunt the advantage.","feed_headline":"Motion-only pseudo labels beat supervised object discovery","feed_subtitle":"From static-camera frames, pure flow clustering lets MR-DINOSAUR top DIOD on TRI-PD and KITTI with no labels.","key_machinery":"The load-bearing mechanism is an unsupervised pseudo-label generator that first labels a frame pair as quasi-static when the average optical-flow magnitude in most image corners is below a threshold, then thresholds flow magnitude to get a foreground mask, extracts connected components, and splits components that contain flow-gradient discontinuities using density-based clustering on flow magnitude, angle, and pixel position. The second mechanism is a two-stage training scheme: stage 1 applies Hungarian matching and a weighted binary cross-entropy loss to pull DINOSAUR's alpha masks onto the pseudo instance masks, and stage 2 trains a slot-deactivation MLP with a foreground/background loss (negative log-likelihood plus background regularization) and a similarity-based drop loss that ignores unmatched slots whose representations resemble matched foreground slots, so static objects are not erased. Together these convert motion-only cues into a foreground/background decision for every slot.","core_discovery":"On the paper's own terms, the central discovery is that the hard part of multi-object discovery is not the object-centric model but the label signal: once high-quality instance masks are obtained without supervision—by exploiting frames where the background is static and flow is caused only by moving objects—a standard slot-attention model such as DINOSAUR can be refined into a state-of-the-art object segmenter. The refinement happens in two stages: a weighted binary cross-entropy loss matches slot alpha masks to the pseudo instance masks and sharpens object boundaries, then a slot-deactivation module learns to switch off background slots, with a similarity-based drop loss that keeps the model from punishing slots that plausibly represent static objects. The paper reports that this fully unsupervised system improves $F_1$ by 6.6 points over DIOD on TRI-PD and 11.8 points on KITTI, and that its own pseudo labels score substantially higher than the supervised TSAM labels used by prior work (15.4 vs 8.9 $F_1$ on the same subset).","pith_inferences":["The paper leaves implicit that its quasi-static frame filter is a data-selection prior: coupling it with estimated ego-motion or inertial cues could extend the pipeline to handheld, drone, or always-moving cameras while keeping the same pseudo-label generator.","A testable consequence of the similarity-based drop loss is a precision-recall trade-off controlled by the slot-similarity threshold; sweeping that threshold on TRI-PD would show whether the module mainly suppresses false positives or also sacrifices genuinely static objects.","Because pseudo labels contain only moving objects, the method's 'object' category is motion-defined; connecting it to image-level object-centricity signals could yield a system that discovers static objects as reliably as dynamic ones, going beyond what the paper evaluates.","The two-stage refinement—first sharpening masks, then learning which slots are objects—could transfer directly to other slot-attention backbones or to weakly supervised panoptic segmentation, where background slots are rarely labeled."],"forward_implications":["Fully unsupervised multi-object discovery is achievable on real driving scenes: on KITTI, MR-DINOSAUR reaches $F_1$ 35.0, beating DIOD without any synthetic pretraining (DIOD drops to 14.1 when retrained on KITTI only).","Motion-derived pseudo labels from quasi-static frames are enough to refine a pre-trained object-centric model; the paper's own pseudo labels outperform the supervised TSAM labels used by earlier methods on the same frames (15.4 vs 8.9 $F_1$).","The similarity-based drop loss lets the model go beyond the moving objects seen in pseudo labels, so static objects of the same category can still be discovered rather than being forced into the background.","The method is agnostic to the base object-centric learner, so future improvements in slot-attention or object-centric backbones can be plugged in to raise accuracy further.","The refinement stage needs only a fraction of the training data (about 13,280 of 94,200 TRI-PD frames) and modest compute, with stage 1 taking about 11 hours and stage 2 about 40 minutes on a single GPU."],"supporting_citations":[{"why":"Supplies the base slot-attention object-centric model whose alpha masks and slot representations are refined.","marker":"[56]"},{"why":"Defines the multi-object discovery task, provides the TRI-PD dataset and evaluation protocol, and introduces the supervised TSAM pseudo-label pipeline that prior work relies on.","marker":"[4]"},{"why":"Provides the unsupervised optical flow used both to identify quasi-static frames and to generate pseudo instance masks.","marker":"[60]"},{"why":"Provides the density-based clustering used to split connected components into separate instance pseudo masks.","marker":"[12]"},{"why":"Defines the slot-attention mechanism inside DINOSAUR whose masks the paper refines.","marker":"[46]"},{"why":"Supplies the self-supervised DINOv2 features that DINOSAUR reconstructs and that the paper uses as the visual backbone.","marker":"[50]"},{"why":"Is the previous state of the art that MR-DINOSAUR compares against, using self-distillation and supervised pseudo labels.","marker":"[35]"},{"why":"Is the supervised motion-segmentation model whose pseudo labels are used by prior methods and serve as the comparison baseline for the paper's own pseudo labels.","marker":"[17]"}],"fun_headline_variants":["Motion-only pseudo labels refine DINOSAUR to SOTA MOD","Unsupervised object discovery via static-frame flow clustering","MR-DINOSAUR: No labels, just motion, beats supervised DIOD","Flow from static frames yields high-quality pseudo masks for MOD","Why object discovery hinges on label signal, not the model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach presumes that training videos contain frames in which the camera is perfectly still; when camera motion is constant, as in the MOVI-E dataset, the motion clues lose their meaning and the method's segmentation quality falls below the comparison method.","fun_headline_variants_meta":{"raw":{"variants":["Motion-only pseudo labels refine DINOSAUR to SOTA MOD","Unsupervised object discovery via static-frame flow clustering","MR-DINOSAUR: No labels, just motion, beats supervised DIOD","Flow from static frames yields high-quality pseudo masks for MOD","Why object discovery hinges on label signal, not the model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1417,"prompt_tokens":950,"completion_tokens":467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":566,"tokens_out":467,"duration_ms":4181,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:36:15.411403+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MR-DINOSAUR on a video dataset where every frame has camera motion, once with the still-frame filter and once with it turned off; if the two versions perform equally, the static-camera assumption is not load-bearing, whereas the paper's own results on the constantly-moving MOVI-E dataset suggest the filtered version should win.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base slot-attention object-centric model whose alpha masks and slot representations are refined."},{"cited_title":"o rkay Aydemir, Weidi Xie, and Fatma G \\","cited_arxiv_id":null,"evidence_quote":"Defines the multi-object discovery task, provides the TRI-PD dataset and evaluation protocol, and introduces the supervised TSAM pseudo-label pipeline that prior work relies on."},{"cited_title":"SF2SE3: C lustering scene flow into SE(3) -motions via proposal and selection","cited_arxiv_id":null,"evidence_quote":"Provides the unsupervised optical flow used both to identify quasi-static frames and to generate pseudo instance masks."},{"cited_title":"Cascade R-CNN : H igh quality object detection and instance segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the density-based clustering used to split connected components into separate instance pseudo masks."},{"cited_title":"Space: Unsupervised object-oriented scene representation via spatial attention and decomposition","cited_arxiv_id":null,"evidence_quote":"Defines the slot-attention mechanism inside DINOSAUR whose masks the paper refines."},{"cited_title":"Higher order motion models and spectral clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised DINOv2 features that DINOSAUR reconstructs and that the paper uses as the visual backbone."},{"cited_title":"The background also matters: B ackground-aware motion-guided objects discovery","cited_arxiv_id":null,"evidence_quote":"Is the previous state of the art that MR-DINOSAUR compares against, using self-distillation and supervised pseudo labels."},{"cited_title":"Motion competition: A variational approach to piecewise parametric motion segmentation","cited_arxiv_id":null,"evidence_quote":"Is the supervised motion-segmentation model whose pseudo labels are used by prior methods and serve as the comparison baseline for the paper's own pseudo labels."}],"review_version":2}