{"id":"f1348a64-7a73-493e-9f5f-db979a8c3230","arxiv_id":"2506.07364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Multiple Object Stitching improves self-supervised representations by training on stitched multi-object images and reports gains on ImageNet, CIFAR, and COCO.","lead":"This paper introduces Multiple Object Stitching (MOS), a self-supervised training trick that stitches small cropped views from different images into artificial multi-object scenes and trains a vision transformer to align each stitched scene with its component views. Readers outside the subfield should care because the method claims to improve representations for images containing many objects, which matter for detection and segmentation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 omits iBOT from the COCO comparison, so the claimed SOTA on multi-object transfer is not yet established.","rationale":"The reader's identified weakest assumption, ViT insensitivity to stitching-boundary artifacts, is a genuine mechanistic concern, but it is less load-bearing than the benchmark gap because the COCO transfer result is itself a direct empirical test of whether stitched-image pretraining transfers to natural multi-object scenes. The missing iBOT baseline is instead a direct threat to the paper's strongest sentence: the claim of significantly outperforming previous SOTA on multi-object images rests entirely on Table 3, and the strongest prior method from the paper's own Table 1 is not included there. If iBOT turns out comparable on COCO, the headline overstates the contribution; if iBOT is clearly worse, the omission is harmless but should still be documented. Either way, the check is cheap and decisive. The reader's CONDITIONAL verdict already reflects the need for additional verification, so this stress-test does not change the recommended verdict.","tokens_in":20213,"tokens_out":12139,"duration_ms":154469,"concrete_test":"Obtain the released iBOT ViT-S/16 ImageNet-1K checkpoint and evaluate it with the exact Mask R-CNN fine-tuning protocol used for Table 3, including the same detector configuration, schedule, and input scale. Record APbb and APmk and compare them with MOS's 45.6 APbb and 40.6 APmk. If iBOT is within 0.5 AP or higher, the claimed 'significantly outperforming previous SOTA on multi-object images' is not established; the authors should add iBOT to Table 3 and report per-seed variance over at least three seeds before the claim is accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MOS 'significantly outperforms previous state-of-the-art methods on both single object centric images and multi-object ones.' The single-object evidence appears in Tables 1 and 2 and includes iBOT. The multi-object evidence is Table 3, the COCO detection and segmentation evaluation, which lists only MoCo v3, DINO, SelfPatch, and ADCLR as ViT-S/16 baselines. iBOT, which the paper itself treats as the strongest prior in Table 1, is absent from Table 3 and from the COCO evaluation entirely. Without that comparison, the multi-object half of the SOTA claim is unsupported: MOS beats the best included COCO baseline, ADCLR, by 1.3 APbb and 0.9 APmk, but no error bars or multiple seeds are reported, so the significance of even that margin is unquantified. If iBOT's Mask R-CNN numbers are close to or above 45.6 APbb and 40.6 APmk under the same protocol, the 'significantly outperforming' statement fails. This is a benchmark-completeness and significance issue, not an internal contradiction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Multiple Object Stitching (MOS), an unsupervised representation-learning method that synthesizes multi-object images by stitching randomly cropped views from single-object-centric images. It introduces three contrastive objectives: multiple-to-single (Eq. 11), multiple-to-multiple (Eq. 12), and single-to-single (Eq. 13), trained with a MoCo-v3-style base/momentum encoder. Experiments report ViT-S/16 and ViT-B/16 results on ImageNet-1K, ViT-T/S/B on CIFAR-10/100, and transfer to COCO object detection and instance segmentation with Mask R-CNN, claiming state-of-the-art performance in both single-object and multi-object settings.","tokens_in":20431,"tokens_out":8504,"duration_ms":92179,"significance":"If the reported results are reproducible, MOS is an attractive and simple contribution: it requires no region proposals, no saliency maps, no additional annotations, and it provides known correspondences among synthetic multi-object views. The paper ships code, reports ablations isolating each loss component, and evaluates across three dataset families and two downstream task types; the proposed losses are easy to implement in existing MoCo/DINO-style frameworks. The main uncertainty is whether the empirical SOTA claim survives a complete baseline set (notably iBOT on COCO) and significance testing, and whether the ViT seam-insensitivity premise holds.","major_comments":[{"comment":"Table 3 omits iBOT from the COCO object-detection and instance-segmentation comparison, even though Section 4.2.1 and Tables 1-2 treat iBOT as the strongest prior. The abstract's claim of 'significantly outperforming previous state-of-the-art methods on both single object centric images and multi-object ones' is therefore not supported for the multi-object half: MOS improves over the best included ViT-S/16 baseline (ADCLR) by only 1.3 APbb and 0.9 APmk, and it is unknown whether iBOT, under the same Mask R-CNN protocol, lies above or below those numbers. Please add iBOT to Table 3 or explicitly restrict the COCO claim to the baselines included.","section":"Sec. 4.2.3, Table 3"},{"comment":"No repeated runs, error bars, or significance tests are reported. The COCO margins are small in absolute terms, and the ablations in Tables 4 and 5 show differences of comparable magnitude (e.g., adding Lm2m changes APbb from 44.3 to 45.6), so 'significantly outperforming' and 'the fusion ... significantly improves' are not statistically supported. Please report mean and standard deviation over at least three seeds for the main results and ablations, or soften the significance claims accordingly.","section":"Sec. 4.2.3, Tables 3-5"},{"comment":"The load-bearing premise that 'Vision Transformer architecture is less sensitive to artificiality produced by the boundary of image stitching' is asserted without a citation, a controlled experiment, or a quantitative check. If the model instead latches onto seam boundaries or grid layouts, the multiple-to-single and multiple-to-multiple losses could be minimized by layout artifacts rather than by object semantics, weakening the claimed mechanism behind the COCO gains. Please provide evidence, for example by comparing hard-boundary stitching with soft-blended boundaries, by probing attention at seam locations, or by benchmarking on natural multi-object crops.","section":"Sec. 1 and Sec. 5"},{"comment":"The method calls each stitched sub-view an 'object', but those sub-views are random crops of a single-object-centric image and may contain background, partial objects, or several object fragments. Thus the 'object correspondences' are really crop correspondences, and the loss in Eq. (11) treats every stitched crop as a positive object regardless of its content. Since the central motivation is object-level contrast, please quantify how often the stitched views actually correspond to foreground objects (e.g., with a detector or saliency oracle) or revise the terminology and motivation.","section":"Sec. 3.1, Eq. (11)"}],"minor_comments":[{"comment":"The sentence 'our method outperforms state-of-the-art iBOT by 1.7% and 1.8% under linear and finetune protocol, respectively' names a finetune comparison for iBOT, but Table 1 lists no finetune accuracy for iBOT; the 1.8% difference matches the kNN column. Please correct the protocol name.","section":"Sec. 4.2.1"},{"comment":"In addition to the missing error bars, the 'NA' entry for the configuration with only Lm2m (failure to converge) would benefit from a brief explanation in the appendix, since it is the only non-convergent configuration and the main text only gives a partial interpretation.","section":"Sec. 4.3.1, Table 4"},{"comment":"The t-SNE and self-attention visualizations are qualitative and depend on hand-picked examples; they should be framed as illustrations rather than as evidence supporting the main quantitative claims.","section":"Appendix F/G"},{"comment":"The notation around the stitching indices is dense and mixes r, s, and r^2; for example Eq. (7) is hard to parse without the appendix, and the relationship between the u(i,j) used near Eq. (5) and the 1-D indices in Eq. (7) should be spelled out in the main text.","section":"Sec. 3.1, Eqs. (1)-(10)"}],"recommendation":"major_revision","confidential_remarks":"The omission of iBOT from Table 3 is the main factual gap in an otherwise broad empirical study. If the authors can add the iBOT COCO numbers under the same protocol and provide error bars, the paper would meet the journal bar. The stitching-boundary premise also needs direct evidence, as the mechanism story depends on it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent empirical contribution. The core idea — stitching r^2 single-object views into synthetic multi-object images and deriving three contrastive losses that exploit the known object correspondences — is genuinely new, and the ablations in Table 4 give real support for each loss component. The evaluation is broad: ImageNet, CIFAR, and COCO detection/segmentation, all with ViT backbones. There is no circularity: the correspondences come from the stitching procedure, not from model outputs, and all downstream evaluations are held out. The authors also include honest discussion of the potential domain gap and pretraining curves, which I appreciate.\n\nThe soft spots are real but addressable. First, the stress-test note is right: Table 3, the only multi-object transfer evidence, omits iBOT, which the paper itself treats as the strongest prior in Tables 1 and 2. Without iBOT on the same COCO Mask R-CNN protocol, the claim of \"significantly outperforming previous state-of-the-art on multi-object images\" is unsupported. The margin over the best included baseline (ADCLR) is 1.3 APbb and 0.9 APmk, but there are no repeated seeds or error bars anywhere in the paper, so even that margin is unquantified. Second, the load-bearing premise that ViT is insensitive to stitching-boundary artifacts is asserted in the introduction and partially repeated in the discussion, but never directly tested (e.g., by comparing stitching with soft blending or by evaluating seam artifacts). That is more of a missing sanity check than a fatal flaw, because the downstream COCO gains suggest the approach works in practice. Third, the writing has rough spots — the appendix contains some garbled font sequences and the formulas for the m2m targets are hard to follow — but these are minor.\n\nThe central argument, taken as an engineering advance in SSL, holds up. The method is simple, plausible, and broadly evaluated. Who is it for? Researchers working on self-supervised pretraining for dense prediction tasks, or anyone who wants a cheap way to inject multi-object supervision without labels. It deserves a serious referee, not a desk reject.\n\nMy recommendation: send it to peer review, but require the authors to add iBOT (and ideally DINO with multi-crop) to the COCO comparison, report results over at least three seeds with standard deviations, and include a small experiment probing stitching-boundary sensitivity. With those additions, the paper would be much stronger and the main claim credible.","headline":"A solid empirical SSL paper with a genuinely new stitching objective, but the multi-object SOTA claim is not yet backed up because the COCO comparison omits iBOT and there are no error bars.","tokens_in":20959,"tokens_out":1167,"would_cite":true,"duration_ms":13109,"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":"Stitching known single-object images into synthetic multi-object composites gives contrastive learning object-level correspondences without labels, and unsupervised representations that top single-object and multi-object benchmarks.","keywords":["unsupervised representation learning","contrastive learning","multiple object stitching","multi-object images","vision transformer","self-supervised learning","object detection","instance segmentation"],"falsifier":"Run the identical three-loss pretraining on stitched images whose seams are made conspicuous, for example by drawing bright grid lines or leaving one- or two-pixel gaps between tiles. If ImageNet linear-probe accuracy and COCO transfer AP stay about the same, seam insensitivity is not what carries the result; if they drop sharply, it is. As a second check, permute the tile-to-source labels during pretraining so $y_{\\mathrm{m2s}}$ no longer matches the tiles; if performance is unchanged, the exact object correspondence is not the active ingredient.","tokens_in":20002,"feed_emoji":"🧩","tokens_out":16526,"duration_ms":144102,"temperature":0.7,"pith_summary":"Contrastive representation learning stumbles on images that contain several objects: random crops from one image can hold different objects, so the method pairs views that do not actually match. The paper sidesteps this by never letting randomness define similarity: it stitches small views of known single-object images into synthetic multi-object images, so every object in a stitched image has an exact, known source. Contrasting a stitched image against its constituent single-object views teaches the model to represent each object rather than the image as a whole. The reported results — leading accuracy on ImageNet and CIFAR plus top detection and segmentation transfer on COCO — support the claim that this simple, label-free trick yields object-level representations without annotations.","feed_headline":"Stitching beats prior self-supervised vision models on ImageNet, COCO","feed_subtitle":"A label-free method learns object-level features by stitching single-object images into synthetic multi-object scenes.","key_machinery":"The carrier of the argument is the Multiple Object Stitching (MOS) operation: a deterministic rearrangement of small augmented views from a batch of single-object images into grid composites, together with the index bookkeeping that makes every object's source known. Each image $x_i$ is transformed into $r^2$ views, optionally at a smaller inner scale $s$, and the flattened view sequence is permuted by $q = (t + (t \\bmod r^2)\\cdot r^2) \\bmod T$ with $T = N r^2$, which mixes views from different batch samples into a stitched image $I_i$ containing $r^2$ objects. The same rule generates the label maps $y_{\\mathrm{m2s}}$ (which source image each tile came from) and $y_{\\mathrm{m2m}}$ with overlap weights $\\omega_{\\mathrm{m2m}}$ (how many objects two stitched images share), so the contrastive targets are exact rather than inferred. Three losses run on this bookkeeping — multiple-to-single $\\mathcal{L}_{\\mathrm{m2s}}$, multiple-to-multiple $\\mathcal{L}_{\\mathrm{m2m}}$, and single-to-single $\\mathcal{L}_{\\mathrm{s2s}}$ applied to natural views to close the synthetic-to-natural gap. The second load-bearing piece is the premise that Vision Transformers, which read images as patch sequences, are insensitive enough to seam boundaries for stitched grids to stand in for natural multi-object scenes.","core_discovery":"The central claim is that synthetic multi-object images with predetermined object identities supply exact object-level correspondences for contrastive learning, eliminating the semantic inconsistency that image-level contrast suffers on natural multi-object images. The authors build each multi-object image by tiling small, scale-varied views drawn from a batch of single-object images; because the tile order follows a fixed cyclic index rule, the labels linking any stitched image to its source images are known exactly. Three contrastive objectives shape the representation: multiple-to-single (the stitched image against each source object view), multiple-to-multiple (two stitched images, weighted by how many objects they share), and single-to-single (natural views, kept to close the domain gap from stitching). On Vision Transformer backbones the paper reports 83.5% finetune, 77.9% linear, and 74.2% kNN accuracy on ImageNet-1K; 98.3%, 96.3%, and 95.1% on CIFAR10; 86.1%, 78.5%, and 73.5% on CIFAR100; and 45.6 box AP with 40.6 mask AP on COCO, taken to show that the model captures per-object detail without sacrificing single-object classification.","pith_inferences":["The exact-correspondence mechanism is testable against a simpler hypothesis: if the same three losses run with tiles assigned to random, incorrect source labels, the correspondence signal is destroyed, and if performance does not drop the paper's mechanism is not what drives the gains; the paper does not run this control.","The same known-composition trick should transfer to any domain whose units can be deterministically assembled — the paper announces plans for audio and video — so the stitching objective could generalize beyond images without new machinery.","The COCO evaluation uses only Mask R-CNN on a ViT-S/16 backbone, so whether the features are genuinely object-level in a backbone-agnostic sense remains open; a convolutional detector or a dense pixel head would give a sharper test of that claim."],"forward_implications":["The representation transfers to dense prediction: a MOS-pretrained ViT-S/16 under Mask R-CNN reaches 45.6 box AP and 40.6 mask AP on COCO, above prior contrastive methods aimed at single-object images and above earlier region- and patch-level multi-object methods.","The multiple-to-single loss carries most of the benefit: ablations show that adding it to the single-to-single baseline raises CIFAR100 linear accuracy by 9.1 points and kNN accuracy by 7.8 points, with clear gains on ImageNet and COCO as well.","The multiple-to-multiple loss alone fails to converge, and the full three-term objective beats every subset, so the single-to-single term anchors training against the synthetic-to-natural gap.","Multi-scale stitching is what pays off on multi-object scenes: sampling tile counts and inner scales from $\\{1,2\\}$ adds 3.3 box AP on COCO over single-scale stitching, while extending the range to $\\{1,2,3\\}$ gives no further gain.","Because object correspondences come free of annotation, the method needs no region proposals or saliency maps, removing the incorrect-proposal failure mode of earlier multi-object contrastive approaches."],"supporting_citations":[{"why":"Supplies the base-and-momentum encoder framework, projection and prediction heads, and the training recipe that MOS runs inside.","marker":"(Chen et al., 2021)"},{"why":"Defines the Vision Transformer backbone whose patch-based processing is the stated reason stitching seams are tolerable.","marker":"(Dosovitskiy et al., 2021)"},{"why":"The state-of-the-art contrastive baseline that MOS must beat; linear, kNN, finetune, and attention-map comparisons are measured against it.","marker":"(Zhou et al., 2022)"},{"why":"The main self-distillation baseline and source of the multi-crop strategy that sets the fair-comparison bar for MOS.","marker":"(Caron et al., 2021)"},{"why":"Prior work modeling inter-instance similarity by mixing images; MOS extends this mixing idea from image patches to object-level stitching.","marker":"(Shen et al., 2023)"},{"why":"Provides the stop-gradient operator used on all projection representations to prevent collapse.","marker":"(Grill et al., 2020)"},{"why":"Defines the Mask R-CNN detector and instance-segmentation head used to measure transfer to multi-object scenes on COCO.","marker":"(He et al., 2017)"},{"why":"The multi-object benchmark whose detection and segmentation AP evaluate whether object-level representations transfer.","marker":"(Lin et al., 2014)"},{"why":"The single-object-centric pretraining dataset and classification benchmark for the main reported results.","marker":"(Russakovsky et al., 2015)"}],"fun_headline_variants":["Stitching images to learn object-level features without labels","Synthetic multi-object scenes boost self-supervised vision models","Stitch single-object views for richer unsupervised representations","Object stitching unlocks better features for multi-object images","Self-supervised stitching teaches models object correspondences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Vision Transformers are insensitive enough to stitching seams that synthetic grid composites behave like natural multi-object photographs; if seam artifacts dominate the training signal, the learned representations would not transfer to real multi-object scenes.","fun_headline_variants_meta":{"raw":{"variants":["Stitching images to learn object-level features without labels","Synthetic multi-object scenes boost self-supervised vision models","Stitch single-object views for richer unsupervised representations","Object stitching unlocks better features for multi-object images","Self-supervised stitching teaches models object correspondences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1362,"prompt_tokens":981,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":597,"tokens_out":381,"duration_ms":4818,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:35:59.877985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical three-loss pretraining on stitched images whose seams are made conspicuous, for example by drawing bright grid lines or leaving one- or two-pixel gaps between tiles. If ImageNet linear-probe accuracy and COCO transfer AP stay about the same, seam insensitivity is not what carries the result; if they drop sharply, it is. As a second check, permute the tile-to-source labels during pretraining so $y_{\\mathrm{m2s}}$ no longer matches the tiles; if performance is unchanged, the exact object correspondence is not the active ingredient.","supporting_citations":[{"cited_title":"ibot: Image bert pre-training with online tokenizer","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art contrastive baseline that MOS must beat; linear, kNN, finetune, and attention-map comparisons are measured against it."},{"cited_title":"Emerging properties in self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"The main self-distillation baseline and source of the multi-crop strategy that sets the fair-comparison bar for MOS."},{"cited_title":"Inter-Instance Similarity Modeling for Contrastive Learning","cited_arxiv_id":"2306.12243","evidence_quote":"Prior work modeling inter-instance similarity by mixing images; MOS extends this mixing idea from image patches to object-level stitching."},{"cited_title":"Bootstrap your own latent-a new approach to self-supervised learning","cited_arxiv_id":null,"evidence_quote":"Provides the stop-gradient operator used on all projection representations to prevent collapse."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"The multi-object benchmark whose detection and segmentation AP evaluate whether object-level representations transfer."},{"cited_title":"Berg, and Li Fei-Fei","cited_arxiv_id":null,"evidence_quote":"The single-object-centric pretraining dataset and classification benchmark for the main reported results."}],"review_version":1}