{"id":"bc5a2082-e17e-44c7-8b29-e23982ed2283","arxiv_id":"2412.20047","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-stage head-to-tail training recipe with optional unlabeled pseudo-labels achieves state-of-the-art long-tailed detection on LVIS v1 without external image-level supervision.","lead":"SimLTD splits long-tailed object detection into three training stages (head-class pre-training, tail-class transfer, and a final fine-tuning on a balanced sample) and optionally uses unlabeled images to boost rare-class accuracy. Its core result is that this simple recipe beats prior methods on LVIS v1 without needing external labeled databases like ImageNet.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The semi-supervised SOTA claim rests on pasting rare training instances into unlabeled images (§3.3), but the ablation cited as support in §4.5 is absent; without a paste/no-paste comparison, the APr gains may reflect memorization rather than transfer.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the pasted rare instances may inflate tail-class adaptation through synthetic over-representation rather than genuine transfer. My stress test sharpens this by locating the specific omission: the paper promises an ablation in §4.5 that is not actually reported there. Without that ablation, the semi-supervised component that enables tail pseudo-labels is unsupported. I also considered the reader's other flagged issues: k=30 is tuned on the validation set used for the final numbers, and the abstract overclaims 'record results' when Table 2 shows RichSem Swin-L at 52.0 mAP vs SimLTD's 51.5 in the semi-supervised row. These are real but secondary: the k=30 effect is a single smooth hyperparameter, and the overclaim is a wording issue. The paste ablation is more load-bearing because the semi-supervised SOTA is the headline contribution and the mechanism is unvalidated. A controlled paste/no-paste comparison would settle it; the paper should also release code to make the numbers independently checkable. Given the substantial fully supervised results and the plausibility of the framework, conditional acceptance is appropriate until this ablation and code are provided.","tokens_in":14638,"tokens_out":5237,"duration_ms":59330,"concrete_test":"Run the semi-supervised Step 2 training exactly as described in §3.3, with the only change being toggling the rare-instance pasting on and off, keeping all hyperparameters and the COCO-unlabeled2017 set fixed, then fine-tune on Dk=30 and report LVIS v1 validation APr and mAP. If the no-paste model retains the reported APr gain over the supervised baseline, the concern is moot; if the gain largely disappears, the semi-supervised record is an artifact of pasting labeled rare instances rather than of pseudo-labeling unlabeled data. A second, sharper check: hold out a subset of rare-class training images from the paste pool, paste only the remaining instances, and measure APr separately on the held-out and seen instances to test memorization directly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Step 2 (§3.3) introduces a new procedure: rare instances from the labeled training set are pasted into unlabeled images so that the teacher model sees them and produces pseudo-labels for tail classes. The paper states that 'our ablation experiments in §4.5 reveal that it is surprisingly helpful,' but §4.5 contains no ablation that isolates this pasting component. Table 4 ablates RFS, Copy-Paste, and multi-stage training in the fully supervised setting; Figure 8 ablates tail initialization and the number of shots k. Neither directly toggles the rare-instance pasting in the semi-supervised pipeline. This is a missing verification of a load-bearing component. The pasted objects carry ground-truth boxes and class labels from the LVIS training set, so the pseudo-label loss is effectively supervised by exact rare training instances, repeated many times across diverse synthetic backgrounds. Since LVIS rare classes have at most 10 training images, the resulting APr gain on validation could be driven by memorization of those specific instances rather than by improved generalization from unlabeled data. The central claim of new record semi-supervised results on LVIS therefore rests on an unverified and potentially artifactual augmentation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SimLTD, a three-stage pipeline for long-tailed object detection: (1) pre-training on head classes, (2) transfer learning on tail classes by fine-tuning only the classifier and regressor, and (3) fine-tuning on a sampled set Dk of k instances per class. The method optionally uses unlabeled images via pseudo-labeling and introduces a procedure that pastes rare training instances into unlabeled images to increase pseudo-labels for tail classes. Experiments on LVIS v1 report strong mAPbox and APr across Faster R-CNN, Deformable DETR, and DINO with ResNet/Swin backbones, including semi-supervised settings with COCO-unlabeled and Objects365-unlabeled images, plus an ancillary Objects365 evaluation. The paper claims new state-of-the-art results on LVIS v1 without external image-level labels.","tokens_in":14896,"tokens_out":7853,"duration_ms":76908,"significance":"If the results hold, SimLTD is significant because it demonstrates that a simple multi-stage head-to-tail transfer with unlabeled pseudo-labels can compete with or beat methods that rely on CLIP and roughly 1.5M labeled ImageNet images, which would broaden the applicability of long-tailed detection to domains without large object-centric databases. The paper is also valuable for its systematic comparisons across detectors and backbones, its APFixed comparison to CascadeMatch, and the Objects365 generalization check. However, the record claims currently rest on an unverified pasting component, point estimates without variance, and hyperparameters selected on the evaluation set, so the significance is conditional on completing the missing ablation and qualifying the claims.","major_comments":[{"comment":"Section 3.3 introduces rare-instance pasting and justifies it by saying that 'our ablation experiments in §4.5 reveal that it is surprisingly helpful,' but Section 4.5 contains no such ablation. Table 4 toggles RFS, Copy-Paste, and multi-stage training in the fully supervised setting, and Figure 8 toggles tail initialization and the number of shots k; neither isolates the paste/no-paste decision in the semi-supervised pipeline. Because the pasted objects are rare training instances with ground-truth boxes, the pseudo-label loss in Step 2 is effectively supervised by those exact exemplars repeated over many backgrounds, so the reported APr gain could reflect memorization of specific training instances rather than transfer from unlabeled data. Please add a paste/no-paste ablation in the semi-supervised setting and, ideally, report pseudo-label statistics such as the number of rare-class pseudo-boxes and their precision to show that the augmented images teach transferable features rather than unrealistic image statistics.","section":"Section 3.3 and Section 4.5"},{"comment":"The claim of 'new record results' on LVIS v1 is too strong as stated. In the semi-supervised Swin-L row, SimLTD MixPL with COCO-unlabeled reaches 51.5 mAPbox / 45.0 APr, while RichSem with ImageNet-21K and CLIP reaches 52.0 / 50.2, so this configuration is not a record. In the supervised Swin-L row, the margin over RichSem is only 0.1 mAPbox (49.8 vs 49.7). Since Section 4.1 states that Dk is sampled three times, please report the variance or standard deviation for the main results and either soften the 'new record' claim or restrict it to the configurations where the evidence actually supports it.","section":"Table 2 and Abstract"},{"comment":"The fine-tuning shot count k = 30 is selected by examining the 'knee in the curve' on the same LVIS v1 validation set used to report all final numbers in Table 2. This makes the reported metrics optimistically biased, which is particularly problematic for a paper whose central claim is a new record by small margins. Please report results on a held-out split, or at least use a validation split only for hyperparameter selection, or include a sensitivity analysis showing that the qualitative conclusions are stable across reasonable choices of k.","section":"Section 4.5, Figure 8"}],"minor_comments":[{"comment":"The phrase 'k = 30000 shots' is inconsistent with the per-class definition of k introduced in Section 3; with 365 classes this would imply a very large number of per-class samples. Please clarify whether k is per class or total, and how this value relates to the 30-shot choice on LVIS.","section":"Section 4.3, Objects365"},{"comment":"The caption says 'The model is updated only during pre-training,' which contradicts Sections 3.3 and 3.4 where the classifier and regressor are updated in Steps 2 and 3. Please clarify that the backbone or encoder-decoder is frozen after Step 1, while the detector head is updated later.","section":"Figure 3 caption"},{"comment":"The statement that the semi-supervised methods 'inherit all hyper-parameters originally tuned on the COCO dataset' leaves the pseudo-label loss weight alpha from Equation (2) and the teacher-update schedule unspecified. Please provide these values or point to the exact configuration in the released code.","section":"Section 4.2"},{"comment":"The phrase 'without extra image labels' should be qualified because the Swin backbones are pre-trained on ImageNet-22K, which uses image-level labels. The claim is understandable as 'without additional training-image labels beyond standard backbone pre-training,' but the current wording is easily misread.","section":"Abstract and Section 4.3"},{"comment":"The annotations '+25%' and '+19%' are not defined; please specify whether they are relative or absolute AP changes and over which baseline they are computed.","section":"Figure 8, left panel"}],"recommendation":"major_revision","confidential_remarks":"The missing paste ablation is the main risk: it is load-bearing for the semi-supervised SOTA claim but is straightforward to add, so I recommend major revision rather than rejection. The overbroad record claim and the validation-set selection of k should also be addressed. I do not see a circularity problem in the learning procedure itself; the concern is evaluation leakage through hyperparameter selection and the absence of variance estimates."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. SimLTD is a three-stage head-to-tail transfer recipe: pretrain on head classes, transfer to tail classes, then fine-tune on a balanced 30-shot replay set. The supervised results are the real news. With a plain Faster R-CNN backbone it beats the prior no-external-data state of the art by clear margins, and the ablation table actually earns the result: multi-stage training plus RFS and copy-paste gives +5 APr over the single-stage baseline. The architecture sweep (R101, R50, Swin, DINO) is useful, and the Objects365 transfer check is a nice generalization datapoint.\n\nThe semi-supervised claim is where I get cautious. The paper says pasting rare instances into unlabeled images is \"surprisingly helpful\" and points to §4.5, but §4.5 does not isolate that component. Table 4 tests RFS/copy-paste/multi-stage in the fully supervised setting; Figure 8 tests transfer initialization and shot count. Neither toggles the rare-paste step in the semi-supervised pipeline. Given that LVIS rare classes have at most 10 training images, the pseudo-label loss on pasted instances could partly reflect memorization of those exact boxes. The stress-test note is right about this gap, and I don't see it addressed anywhere in the text. That missing ablation matters because the abstract claims \"new record results across both supervised and semi-supervised settings,\" and in the semi-supervised table RichSem Swin-L still holds the top mAP and APr on several rows. The comparison is favorable to SimLTD only if you weight the no-external-data advantage heavily; that's a legitimate framing but the abstract should say it.\n\nSmaller issues: k=30 is tuned on the same validation set used for final numbers, there are no error bars in the main table, and the code link is absent even though the implementation section promises it. The single-author benchmark numbers deserve independent checking, but the method is simple enough that reimplementation is feasible.\n\nNet: the supervised contribution is solid and publishable; the semi-supervised contribution has a load-bearing verification hole. A serious referee should engage—this is exactly the kind of paper where a reviewer can ask for the paste/no-paste ablation in the semi-supervised setting, plus variance or at least multiple-seed runs. I'd accept it into review, with the expectation of heavy revision on the semi-supervised claims.\n\nYours,\n[Name]","headline":"A clean three-stage recipe that gets strong LVIS numbers without ImageNet/CLIP, but the semi-supervised headline overreaches and one key ablation is missing.","tokens_in":15408,"tokens_out":634,"would_cite":true,"duration_ms":8567,"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":"SimLTD claims that a simple three-stage head-to-tail transfer, optionally using unlabeled images via pseudo-labeling, achieves new record results on LVIS v1 without any extra image-level labels or pretrained vision-language models.","keywords":["long-tailed object detection","semi-supervised object detection","head-to-tail transfer","pseudo-labeling","LVIS benchmark","class imbalance","few-shot learning","data augmentation"],"falsifier":"Train the semi-supervised pipeline identically but disable the rare-instance pasting in Step 2 (or paste head-class objects instead), and compare rare-class APr on LVIS v1; a non-material drop would show that the pasted-object pseudo-labels, not the multi-stage curriculum, are responsible for the reported gains.","tokens_in":14397,"feed_emoji":"🎯","tokens_out":6322,"duration_ms":59735,"temperature":0.7,"pith_summary":"The paper argues that long-tailed object detection does not need a huge external labeled database. It proposes SimLTD, a three-stage recipe that pre-trains on head classes, transfers to tail classes, then fine-tunes on a balanced sample, optionally using unlabeled images through pseudo-labeling. On the LVIS v1 benchmark, this simple pipeline establishes new record results under both fully supervised and semi-supervised settings, beating methods that rely on vision-language models and millions of extra labeled images. The implication is that easily collected unlabeled images can substitute for costly human annotations in boosting rare-class accuracy.","feed_headline":"Unlabeled images alone now match labeled-ImageNet long-tail detection","feed_subtitle":"SimLTD's three-stage head-to-tail transfer beats CLIP-based rivals on LVIS using only unlabeled images.","key_machinery":"The central mechanism is the three-stage head-to-tail transfer with head-tail class fusion. Stage one trains a detector on the 866 head classes (with repeat-factor sampling and copy-paste). Stage two copies those weights and fine-tunes only the box classifier and regressor on the 337 tail classes, optionally with unlabeled images whose rare instances are augmented by pasting labeled tail objects into diverse scenes so the teacher sees enough tail examples to propose pseudo-labels. Stage three merges the two detectors' classifier and regressor weights via a simple index map (Algorithm 1) and fine-tunes on a 30-shot-per-class replay set to prevent catastrophic forgetting.","core_discovery":"SimLTD's central discovery is that decoupling learning into head pre-training, tail transfer, and balanced fine-tuning, with pseudo-labels on unlabeled images, is enough to match or exceed far more complex systems. The paper shows empirically that stronger COCO-pretrained representations transfer better to LVIS rare classes, then builds a training curriculum that exploits this. Adding unlabeled images improves both head representation learning and tail adaptation, and the method scales across convolutional and transformer detectors. Removing the external labeled-image crutch does not hurt; in the fully supervised setting SimLTD already outperforms those label-hungry competitors on rare-class AP.","pith_inferences":["A natural extension not explored in the paper: the copy-paste of tail instances into unlabeled scenes is a cheap form of synthetic oversampling of rare classes, and similar mechanisms might improve other imbalanced recognition tasks like instance segmentation or open-vocabulary detection.","If the gains hold, the dependence of prior long-tailed methods on large pretrained vision-language models may be a workaround for a weak training curriculum rather than a necessary ingredient, suggesting that stronger curricula could further close the gap with fully supervised data.","One testable extension is to apply the same head-to-tail transfer to bespoke domains like aerial or medical imagery, where the paper argues no equivalent of ImageNet exists but large unlabeled corpora do.","The method's reliance on pseudo-label correctness means that better teacher models or calibrated confidence thresholds should directly improve tail transfer, a prediction that can be checked by ablating the teacher's accuracy."],"forward_implications":["Long-tailed detection can be advanced with unlabeled images alone, removing the requirement of building or licensing a large object-centric labeled corpus.","The same three-stage recipe transfers across detectors and backbones, from Faster R-CNN to DINO with Swin backbones, so practitioners can apply it to their own architectures.","Semi-supervised SimLTD exceeds previous methods that use ImageNet labels and a CLIP classifier on rare-class AP with ResNet backbones, and matches them with larger Swin backbones.","Supervised SimLTD, using no external data at all, already outperforms earlier multi-stage long-tailed detectors such as the three-step smooth-tail method using Deformable DETR.","The 30-shot fine-tuning sweet spot offers a practical rule for balancing head-class preservation and tail-class accuracy."],"supporting_citations":[{"why":"Provides the LVIS v1 dataset, defines the head/tail split, and supplies the evaluation benchmark and metrics.","marker":"[15]"},{"why":"Represents the ImageNet-21K label plus CLIP baseline that SimLTD competes against.","marker":"[49]"},{"why":"A state-of-the-art method using CLIP and ImageNet supervision; serves as the main external-data comparison.","marker":"[27]"},{"why":"A recent three-step multi-stage long-tailed detection method whose strategy SimLTD directly compares with.","marker":"[12]"},{"why":"Early multi-stage learning-to-segment-the-tail work that uses knowledge distillation and is contrasted with the simpler approach.","marker":"[19]"},{"why":"Supplies the SoftER Teacher semi-supervised few-shot detection method used in Step 1.","marker":"[35]"},{"why":"Provides the MixTeacher pseudo-labeling method used for semi-supervised head representation learning.","marker":"[25]"},{"why":"Supplies the MixPL mixup and mosaic pseudo-label method used with transformer-based detectors.","marker":"[7]"},{"why":"A semi-supervised long-tailed detection method using the same unlabeled COCO data; provides a direct comparison.","marker":"[44]"}],"fun_headline_variants":["SimLTD: unlabeled images beat labeled rivals in long-tail detection","Three-stage SimLTD drops ImageNet labels, boosts long-tail detection","No annotation burden: SimLTD matches CLIP-based rivals on LVIS","Long-tail detection without labeled images: SimLTD's simple fix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pasting rare labeled instances into unlabeled images produces pseudo-labels that genuinely improve tail-class adaptation; if those pasted scenes are unrealistic or over-represent the tail categories, the reported transfer gains could be an artifact of the augmentation rather than a real improvement in learning rare-class features.","fun_headline_variants_meta":{"raw":{"variants":["SimLTD: unlabeled images beat labeled rivals in long-tail detection","Three-stage SimLTD drops ImageNet labels, boosts long-tail detection","No annotation burden: SimLTD matches CLIP-based rivals on LVIS","Long-tail detection without labeled images: SimLTD's simple fix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3093,"prompt_tokens":877,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":2137}},"tokens_in":493,"tokens_out":2216,"duration_ms":17858,"temperature":1.0,"reasoning_tokens":2137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:36:32.280157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the semi-supervised pipeline identically but disable the rare-instance pasting in Step 2 (or paste head-class objects instead), and compare rare-class APr on LVIS v1; a non-material drop would show that the pasted-object pseudo-labels, not the multi-stage curriculum, are responsible for the reported gains.","supporting_citations":[{"cited_title":"LVIS: A Dataset for Large V ocabulary Instance Segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the LVIS v1 dataset, defines the head/tail split, and supplies the evaluation benchmark and metrics."},{"cited_title":"Detecting Twenty-Thousand Classes Using Image-Level Supervision","cited_arxiv_id":null,"evidence_quote":"Represents the ImageNet-21K label plus CLIP baseline that SimLTD competes against."},{"cited_title":"Learning from Rich Semantics and Coarse Locations for Long-Tailed Object Detection","cited_arxiv_id":null,"evidence_quote":"A state-of-the-art method using CLIP and ImageNet supervision; serves as the main external-data comparison."},{"cited_title":"Boosting Long-Tailed Object Detection via Step-Wise Learning on Smooth-Tail Data","cited_arxiv_id":null,"evidence_quote":"A recent three-step multi-stage long-tailed detection method whose strategy SimLTD directly compares with."},{"cited_title":"Learning to Segment the Tail","cited_arxiv_id":null,"evidence_quote":"Early multi-stage learning-to-segment-the-tail work that uses knowledge distillation and is contrasted with the simpler approach."},{"cited_title":"LEDetection: A Simple Framework for Semi- Supervised Few-Shot Object Detection","cited_arxiv_id":null,"evidence_quote":"Supplies the SoftER Teacher semi-supervised few-shot detection method used in Step 1."},{"cited_title":"MixTeacher: Mining Promising La- bels with Mixed Scale Teacher for Semi-Supervised Object Detection","cited_arxiv_id":null,"evidence_quote":"Provides the MixTeacher pseudo-labeling method used for semi-supervised head representation learning."},{"cited_title":"Semi-Supervised and Long-Tailed Object Detection with CascadeMatch","cited_arxiv_id":null,"evidence_quote":"A semi-supervised long-tailed detection method using the same unlabeled COCO data; provides a direct comparison."}],"review_version":1}