{"id":"0d732bae-0a41-435c-b858-14766c7c027d","arxiv_id":"2508.20551","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-only contrastive auxiliary branch with a linearly decaying loss weight improves video object detection accuracy by about two mAP points on ImageNet VID without increasing inference cost.","lead":"This paper adds a contrastive learning side-branch to a video object detector, pulling frames of the same video together and pushing different videos apart during training, then discards the branch at test time. The result is a roughly two point mAP gain on ImageNet VID with no added inference cost, reaching 84.0% mAP with ResNet-101 and 85.2% with ResNeXt-101.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central SOTA claim rests on a single unverified comparison to a literature TROI baseline; no code, seeds, or variance are reported.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing point: the entire paper's contribution is a measured improvement over a baseline, but the baseline's provenance is not documented, no code or configuration is released, and all results are single runs without seeds or variance. The ablation in Table 2 is internally consistent and the mechanism is plausible, so this is not a rejection of the method; it is a call for the comparison to be verifiable. A secondary notation issue in the InfoNCE loss and the stated negative-pair count does not change the main verdict, because the empirical claim could still hold even if the text's accounting is sloppy. The appropriate disposition remains conditional acceptance pending a reproducible baseline and seed-variance reporting, matching the reader's conditional verdict.","tokens_in":8495,"tokens_out":6737,"duration_ms":66432,"concrete_test":"Reproduce the TROI baseline by running the same codebase with the CAB branch and DLW disabled, using the exact hyperparameters, data split, and evaluation protocol described in Section 4.1, for at least three random seeds, and report mean and standard deviation together with the CLAB runs under identical seeds. If the reproduced baseline differs from 82.0 mAP by more than about 0.5 mAP, or if the CLAB-minus-TROI gap shrinks below about 1 mAP, then the claimed gain and state-of-the-art status are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the contrastive auxiliary branch and dynamic loss weighting add a consistent ~2.0 mAP gain over TROI at zero inference cost. For that attribution to hold, the TROI baseline (82.0 R101 in Table 1 and Table 2) must have been produced under the exact same data split (VID+DET, FGFA split), training recipe (7 epochs, SGD lr 0.01, batch 16 on two A100 GPUs, one target frame plus 2 support frames), and evaluation protocol (30 support frames, NMS 0.5, shorter side 600) as the CLAB runs. Section 4.1 says 'Except for our CAB module, the model's architecture and hyperparameters follow those proposed by TROI,' but it never states that the baseline was retrained by the authors, and no code, configuration files, seeds, or per-run logs are provided. All reported numbers are single runs. This matters because the 2.0-point gap is exactly the quantity on which the paper's conclusion depends; if the baseline is an older literature value with different training details, or if the validation-set-tuned hyperparameters (temperature tau=0.1, loss weight w=0.005, decay step k=25,000) are overfit to the validation set, the 'state-of-the-art without post-processing' claim could be an artifact of the comparison rather than of the method. A secondary internal inconsistency appears in the positive/negative pair accounting: with N=16 videos and T=3 frames, each anchor has 2 positives and 45 negatives, not '48 negative pairs' as stated in Section 4.1; this suggests the implemented InfoNCE loss may differ from Equation 1, although it is secondary to the baseline attribution issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLAB, a training-only auxiliary-branch method for video object detection. A contrastive InfoNCE loss is applied to intermediate backbone features of video frames, treating frames from the same video as positives and frames from other videos as negatives, and the auxiliary loss weight decays linearly to zero by the midpoint of training (DLW). The method is built on SELSA+TROI and evaluated on ImageNet VID (FGFA split), reporting 84.0 mAP with ResNet-101 and 85.2 mAP with ResNeXt-101, which the authors describe as state-of-the-art for CNN-based models without extra inference cost. Ablations attribute +1.4 mAP to the contrastive auxiliary branch and +0.6 mAP to dynamic loss weighting over a nominal TROI baseline of 82.0 mAP.","tokens_in":8834,"tokens_out":6583,"duration_ms":57736,"significance":"If the reported gains are reproducible, the paper makes a useful and practical contribution: a conceptually simple, inference-cost-free training technique that improves video object detection, with ablations of temperature and loss weight that provide concrete guidance for practitioners. The method is not circular, since performance is measured on an external benchmark, and the reported hyperparameter sensitivity is informative. However, the central claim of a consistent ~2.0 mAP gain rests on a single unverified comparison to a literature baseline and on single-run numbers, so the significance cannot be fully assessed at present.","major_comments":[{"comment":"The baseline 'Temporal ROI Baseline [11]' is reported as 82.0 mAP, but the text does not state whether this number was obtained by the authors' own re-training of TROI under the exact protocol used for CLAB (FGFA split, 7 epochs, batch size 16 on two A100 GPUs, SGD lr 0.01, one target frame plus two support frames, 30 support frames at inference, NMS 0.5, shorter side 600). If 82.0 is taken from the TROI publication, the reported +1.4 and +2.0 gains are not controlled comparisons and could reflect differences in training recipe or evaluation rather than the effect of CAB/DLW. Please state explicitly how the baseline was produced; if it was not retrained in-house, retrain it under identical conditions and report the resulting number.","section":"§4.1, Table 2"},{"comment":"All experimental results are single runs; no seeds, variance estimates, or released code/configurations are provided. The claimed total gain of 2.0 mAP is a difference between two single numbers, and the individual contributions (1.4 and 0.6 mAP) are small relative to typical run-to-run variation in this setting. Please report results from at least three seeds per configuration (mean ± std) and, ideally, release training configuration files and seeds so that the reproducibility of the main claim can be assessed.","section":"§4.3, Tables 2-4"},{"comment":"There are related inconsistencies in the formulation of the contrastive loss. Eq. (1) defines the InfoNCE loss with a single positive sample zi' per anchor, whereas §4.1 states that each image has 2 positive pairs because T=3 frames are sampled per video; if the implemented loss uses both support frames as positives, Eq. (1) must sum over the set of positives in the numerator. In addition, the denominator's indicator 1[k≠i] compares a frame index k with a video index i and should instead exclude all frames from the same video (e.g., 1[v(k)≠i] with v(k) the video of frame k). The negative-pair count '48 negative pairs' is also inconsistent with N=16 videos and T=3 frames, which gives (16−1)×3 = 45 negatives per anchor. Please correct the equation, the indicator notation, and the pair counts so that the method is unambiguous and reproducible.","section":"§3.1, Eq. (1); §4.1"}],"minor_comments":[{"comment":"The word 'constrastive' is a typo and should be 'contrastive' in both the abstract and the introduction.","section":"Abstract, Section 1"},{"comment":"The label 'B frames' near the input tensor is unexplained and conflicts with the notation 'T frames per video'; please use a single consistent notation for the number of frames per video.","section":"Figure 2"},{"comment":"No sources are given for the baseline mAP values of the compared methods; please add a citation for each value or state that the numbers are taken from the corresponding papers.","section":"Table 1"},{"comment":"The sentence 'SELSA and TROI are simplified versions of our approach' is confusing, since CLAB is built upon TROI rather than being a simplification of SELSA/TROI; please rephrase to avoid this inversion.","section":"§4.2"},{"comment":"Eq. (2) defines k=25,000 as the midpoint of training and the text calls it the midpoint; please state the total number of training steps or the per-epoch step count so that the schedule can be verified, since '7 epochs' alone does not determine k.","section":"§3.2, §4.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is simple and the reported direction of the gains is plausible, but the absence of a controlled baseline and of variance information makes the central claim currently unverifiable. If the authors can demonstrate that the TROI baseline was retrained under identical conditions and provide multi-seed results, the paper could be suitable for publication. The comparison table would also benefit from explicit sourcing of baseline numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean, modest engineering contribution: a training-only contrastive auxiliary branch and a linearly decaying loss weight, grafted onto SELSA+TROI, that claims about 2 mAP on ImageNet VID at zero inference cost. The idea is not new—contrastive losses and adaptive auxiliary weights are known—but the specific combination and the reported numbers are new, and the ablation is internally consistent.\n\nWhat the paper does well: the mechanism is simple and plausible, the ablation shows incremental gains (82.0 to 83.4 with CAB, to 84.0 with DLW), and the no-inference-cost claim is real since the branch is discarded at test time. The hyperparameter sweeps for temperature and loss weight are reasonable, and the method beats several strong baselines in the comparison table.\n\nThe soft spots are mostly about reproducibility. The central comparison is against a TROI baseline of 82.0 mAP, but the paper never says whether that baseline was retrained by the authors under the same recipe. It says architecture and hyperparameters follow TROI except for CAB, but no code, configs, seeds, or per-run logs are given. All numbers are single runs, with no error bars or multiple seeds, and the hyperparameters (tau, w, decay step) are tuned on the validation set. These are fixable but real gaps.\n\nThere is also a small internal inconsistency: with N=16 videos and T=3 frames per video, the InfoNCE denominator has (N-1)*T = 45 negatives per anchor, not the '48 negative pairs' stated in Section 4.1. That suggests the implemented loss may differ from Equation 1, or the text has a typo. Either way, it's a detail the authors should clarify.\n\nOverall, the central argument holds up: the method is not circular, the gain is plausible, and the claim is against an external benchmark. The absence of reproducibility details prevents me from calling the exact numbers reliable, but this is a conditional accept situation, not a rejection.\n\nThis paper is for practitioners in video object detection who want a cheap training-time boost. It deserves a serious referee, mostly to verify the baseline and push for code and seeds.","headline":"A simple, credible training-only contrastive auxiliary branch for video object detection, but the headline gain rests on an unreproduced baseline and single-run numbers.","tokens_in":9400,"tokens_out":2200,"would_cite":false,"duration_ms":19211,"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":"Adding a training-only contrastive branch to a video object detector raises its accuracy on ImageNet VID to 84.0% mAP (ResNet-101) and 85.2% mAP (ResNeXt-101) with zero additional inference cost.","keywords":["video object detection","contrastive learning","auxiliary branch","InfoNCE loss","dynamic loss weighting","ImageNet VID","motion blur robustness","temporal feature aggregation"],"falsifier":"Re-run the TROI baseline and CLAB under identical conditions—same code, seeds, hyperparameters, FGFA split, and evaluation script—for at least three seeds. If the average gap is below about 0.5 mAP, or if the variance is comparable to the reported 2.0-point difference, the claim of a consistent gain from the auxiliary branch would be falsified.","tokens_in":8267,"feed_emoji":"🎥","tokens_out":6985,"duration_ms":56825,"temperature":0.7,"pith_summary":"The paper aims to show that video object detection can be improved substantially by adding a contrastive auxiliary branch that is used only during training and discarded at inference. The method, called CLAB, builds on the TROI detector and adds an InfoNCE-style contrastive loss that pulls together frames from the same video and pushes apart frames from different videos, improving robustness to motion blur, occlusion, and rare poses. A dynamic loss weight starts the auxiliary loss at a small value and linearly decays it to zero by mid-training, so the branch shapes the backbone early without disturbing the detection task later. On ImageNet VID, CLAB reaches 84.0% mAP with ResNet-101 and 85.2% mAP with ResNeXt-101, outperforming its TROI baseline by 2.0 and 0.9 points respectively, and the ablation attributes 1.4 points to the contrastive branch and 0.6 points to the weighting strategy.","feed_headline":"Training-only contrastive branch lifts video detection to 84% mAP","feed_subtitle":"Auxiliary InfoNCE branch with decaying loss weight adds up to 2.0 mAP over the TROI baseline at zero inference cost.","key_machinery":"The machinery is CAB (contrastive auxiliary branch), a small trainable head attached to an intermediate backbone feature map: a 3x3 convolution, ReLU, adaptive average pooling, and a two-layer projection to a 128-dimensional vector, followed by the InfoNCE loss with temperature $\\tau=0.1$. This branch is the only addition to the TROI/SELSA detector, and it is removed at inference. The second piece is DLW (dynamic loss weighting), the linearly decaying schedule $w(t)=w\\,\\max(0,1-t/k)$ that lets the auxiliary task dominate early and vanish by step $k=25,000$. Together they shift the backbone toward features that keep frames of the same video close in embedding space, without changing the detector's inference-time architecture or computational cost.","core_discovery":"CLAB's central claim is that a training-only contrastive auxiliary branch can make a video detector's backbone features more resistant to the degradations common in video—motion blur, occlusion, unusual poses—and that this transfers into higher detection accuracy at zero inference cost. The auxiliary branch, attached to an intermediate backbone feature map, projects it to a 128-dimensional embedding and applies the InfoNCE loss with frames from the same video as positives and frames from other videos as negatives. The auxiliary loss is combined with the standard detection losses through a dynamic weight $w(t)=w\\,\\max(0,1-t/k)$ that decays linearly to zero at step $k$, the midpoint of training. On ImageNet VID, CLAB reaches 84.0% mAP with ResNet-101 and 85.2% mAP with ResNeXt-101, exceeding its TROI baseline by 2.0 and 0.9 points respectively and matching or beating published CNN-based video object detectors that use no post-processing.","pith_inferences":["CLAB's principle—a cheap contrastive head whose weight is annealed to zero—could transfer to other video recognition tasks such as action recognition or multi-object tracking, where same-video positives are equally natural and inference cost is a constraint.","The reliance on a single run and the unstated variance across seeds leaves open the possibility that part of the 2.0-point gain is noise; re-running with several seeds and reporting mean and standard deviation would settle how robust the margin is.","If the gain is real, it suggests that backbone representations, rather than temporal aggregation modules, are a bottleneck for robustness to video degradation; that would motivate contrastive pretraining objectives for video detectors more broadly.","The method should combine with post-processing refinements like Seq-NMS, since the gains come from a stronger backbone and do not depend on the scoring stage; combining may push the absolute mAP further."],"forward_implications":["On the ImageNet VID validation set, CLAB with ResNet-101 reports 84.0% mAP and with ResNeXt-101 85.2% mAP, both above published CNN-based detectors evaluated without post-processing.","The method adds zero inference-time cost: CAB and DLW are removed at test time, so the deployed model is identical in compute to TROI while the backbone is trained to be more robust.","The ablation attributes 1.4 mAP of the 2.0-point gain to the contrastive branch and 0.6 mAP to the dynamic loss weighting, indicating that both components contribute.","Because the auxiliary branch is attached to an intermediate feature map and trained with same-video positives, the learned robustness is specific to the detection backbone and does not require optical flow or external memory at inference.","Loss weight and temperature ablations show CLAB stays above the TROI baseline across tested hyperparameter values, suggesting the gain is not limited to a single setting."],"supporting_citations":[{"why":"TROI baseline: supplies the detector architecture, training hyperparameters, and the 82.0 mAP reference point CLAB improves on.","marker":"[11]"},{"why":"SELSA: provides the sequence-level temporal aggregation module that CLAB's detector is built on.","marker":"[27]"},{"why":"Faster R-CNN: supplies the base two-stage detection framework (RPN and ROI pooling) used throughout.","marker":"[21]"},{"why":"SimCLR: source of the projection head design and temperature-based InfoNCE contrastive objective used by CAB.","marker":"[4]"},{"why":"Contrastive Predictive Coding: defines the InfoNCE loss that the auxiliary branch applies to same-video positives and cross-video negatives.","marker":"[19]"},{"why":"MoCo: further informs the contrastive loss formulation and temperature weighting in the auxiliary branch.","marker":"[15]"},{"why":"FGFA: defines the ImageNet VID+DET training split used for all experiments.","marker":"[33]"},{"why":"STSN: provides the support-frame sampling and boundary-repetition rule adopted in training and inference.","marker":"[2]"},{"why":"ImageNet: supplies the VID detection dataset and class definitions used for evaluation.","marker":"[8]"}],"fun_headline_variants":["Auxiliary contrastive branch sharpens video detection to 84% mAP","Training-only contrastive loss makes video detection robust, zero added cost","Contrastive aux branch: better video detection with no inference overhead","Zero-cost training branch boosts video detection to 84%+ mAP","Contrastive auxiliary branch: 84% mAP on video, no inference overhead"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire 2.0-point gain rests on the assumption that the 82.0% mAP TROI baseline was reproduced under exactly the same training recipe, data split, hyperparameters, and evaluation protocol as the CLAB runs; the paper reports one run per configuration with no variance or seed information.","fun_headline_variants_meta":{"raw":{"variants":["Auxiliary contrastive branch sharpens video detection to 84% mAP","Training-only contrastive loss makes video detection robust, zero added cost","Contrastive aux branch: better video detection with no inference overhead","Zero-cost training branch boosts video detection to 84%+ mAP","Contrastive auxiliary branch: 84% mAP on video, no inference overhead"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001036,"raw_usage":{"total_tokens":4372,"prompt_tokens":965,"completion_tokens":3407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":3308}},"tokens_in":581,"tokens_out":3407,"duration_ms":23744,"temperature":1.0,"reasoning_tokens":3308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:42:10.428610+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the TROI baseline and CLAB under identical conditions—same code, seeds, hyperparameters, FGFA split, and evaluation script—for at least three seeds. If the average gap is below about 0.5 mAP, or if the variance is comparable to the reported 2.0-point difference, the claim of a consistent gain from the auxiliary branch would be falsified.","supporting_citations":[{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"TROI baseline: supplies the detector architecture, training hyperparameters, and the 82.0 mAP reference point CLAB improves on."},{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"SELSA: provides the sequence-level temporal aggregation module that CLAB's detector is built on."},{"cited_title":"In: Proceedings of the IEEE international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"FGFA: defines the ImageNet VID+DET training split used for all experiments."},{"cited_title":"In: Proceedings of the European Conference on Computer Vision (ECCV)","cited_arxiv_id":null,"evidence_quote":"STSN: provides the support-frame sampling and boundary-repetition rule adopted in training and inference."}],"review_version":2}