{"id":"3eefc62c-a1de-4e8e-8ab3-360822546863","arxiv_id":"2506.20550","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Multi-frame early fusion with single-frame supervision improves YOLOv7-tiny detection on MOT20Det and a new BOAT360 fisheye dataset.","lead":"This paper feeds a stack of recent video frames into a YOLO detector instead of a single image, labeling only the newest frame. The method improves detection for small models on two benchmarks, including a new fisheye boat dataset the authors contribute.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc selection of frame count and lack of multiple-seed statistics may inflate the reported detection gains.","rationale":"The reader's weakest assumption pointed to the dataset-dependent, post-hoc selection of the number of frames. I agree that this is a weak point, and I sharpen it into a concrete selection-bias concern: with 9+ configurations evaluated on the same test set and no multiple-seed statistics, the best reported numbers may overstate the method's true advantage. However, I do not think this concern alone overturns the paper's plausibility. The 3-frame adjacent configuration improves over the single-frame baseline on both datasets (MOT20Det: 0.841 vs 0.797; BOAT360: 0.539 vs 0.411), and the paper is transparent about the non-monotonic trends. The central mechanism—early fusion of temporally adjacent frames with single-frame supervision—is simple and has face validity. The missing piece is a statistically disciplined evaluation: validation-based hyperparameter selection, multiple seeds, and ideally a shuffled-frame control to confirm that temporal order (not just multi-frame input) drives the gain. My recommended verdict remains CONDITIONAL, consistent with the reader, but with the condition sharpened to require this concrete selection-and-reproducibility check. I give credit for using a public dataset (MOT20Det) and reporting parameter/computation overhead, but the new BOAT360 dataset is not yet released, which adds to the reproducibility concern.","tokens_in":8896,"tokens_out":6542,"duration_ms":84047,"concrete_test":"Use the 1/5 validation split of MOT20Det (and the corresponding split of BOAT360) to evaluate all configurations in Tables III and IV once, then pre-register the single best configuration on validation. Train that configuration from 3 independent seeds on the full training set and report mean ± std mAP@0.5 and mAP@0.5:0.95 on the official test split. If the validation-selected configuration's test gain over the single-frame baseline is less than half the best-in-table gain, or if the 95% confidence interval includes zero, the headline claim is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that multi-frame integration 'improves detection robustness, especially for lightweight models' rests on test-set numbers in Tables III, IV, and IX. The authors evaluate many configurations (n = 1, 3, 5, 7, 9 adjacent; several step-sampled variants; two fusion strategies) and then report the best-performing n for each dataset as evidence. This is a post-hoc selection on the test set, with no validation-based model selection and no error bars or multiple seeds. The risk is concrete: on MOT20Det, 5-frame input is worse than both 3-frame and 7-frame (mAP@0.5: 0.819 vs 0.841 and 0.855), and on BOAT360 the best n is 3, not 7. Such non-monotonicity is consistent with noise or test-set overfitting rather than a stable temporal-context benefit. On BOAT360, the dataset contains only 509 total boxes, so the +12.8 mAP gain for YOLOv7-tiny could easily be a high-variance result. Because the headline improvement is the only quantitative support for the method, the lack of a disciplined selection protocol means the reported effect size—and possibly its existence—is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to improve video object detection by concatenating n consecutive frames along the channel dimension as input to YOLOv7 and YOLOv7-tiny, while supervising only the latest frame of each stack. It compares early fusion with grouped convolution, adjacent with stepped temporal sampling, and evaluates on MOT20Det and a new fisheye BOAT360 dataset. The main empirical claim is that this minimal architectural change improves detection robustness, especially for lightweight models, with negligible parameter and latency overhead.","tokens_in":9104,"tokens_out":7544,"duration_ms":77555,"significance":"If the reported gains are stable, the proposal is attractive because it requires changing only the first convolutional layer and preserves real-time inference on embedded hardware. The paper provides extensive ablations (frame count, sampling step, fusion type, model scale), explicit complexity and speed numbers (Tables VII and VIII), and contributes a new annotated fisheye dataset. It does not, however, provide external comparisons to existing video object detection methods, multi-seed statistics, or a validation-based protocol for choosing the temporal window; these gaps currently limit the strength of the central empirical claim.","major_comments":[{"comment":"The temporal window size is selected on the test set, making the reported gains a form of test-set overfitting. Section IV-C2 states \"Since 7 adjacent frames yielded the best accuracy\" after reporting test numbers, and for BOAT360 the paper presents 3 frames as best because that configuration has the highest mAP in Table IX. The validation split described in Section IV-A is not used for model selection, and no holdout protocol is reported. Because performance is non-monotonic in n (Table III: 5 frames, 0.819, is below both 3 and 7 frames; Table IX: 7 frames, 0.537, is below 3 frames, 0.539), the headline improvement may reflect selection among many configurations. Please select n using validation data only, or report all configurations with confidence intervals so the selection effect can be assessed.","section":"IV-C, Tables III, IV, IX"},{"comment":"The paper never compares against any existing video object detection method. Related work in Section II-B names FGFA, temporal attention, recurrent memory, and YOLOV, but the experiments only contrast the proposed multi-frame models with single-frame YOLO baselines. Without an accuracy/latency comparison against a representative video detector on the same data, the claimed advantage of lightweight multi-frame integration over established video detection approaches is not established.","section":"II-B and IV-C"},{"comment":"All reported metrics come from a single training run. Many differences that support the conclusion are small (e.g., Table VI: YOLOv7 3-frame mAP@0.5:0.95 is 0.520 vs 0.511 for the single-frame baseline; Table III: 7-frame vs 3-frame mAP@0.5 is 0.855 vs 0.841). Without means and standard deviations across multiple seeds, these differences could be run-to-run noise. Please report at least three seeds for the main configurations and, for BOAT360, per-sequence results, since the total annotation count is only 509 boxes (Table II).","section":"IV-C, Tables III-VI, IX"},{"comment":"The BOAT360 dataset is extremely small, with 509 boxes total, so the +12.8 mAP@0.5 gain for YOLOv7-tiny in Table IX may be driven by a handful of frames. The paper should state the number of test images and boxes, report per-sequence metrics, and provide confidence intervals or bootstrap estimates. Without this, the generalization claim based on BOAT360 is fragile.","section":"IV-A, Table II, Table IX"}],"minor_comments":[{"comment":"The inline math notation such as \"{I t}T t=1\" and \"I t−n+1, . . . , It−1, It\" has missing braces and subscripts; please reformat as \\(\\{I_t\\}_{t=1}^T\\) and \\(\\{I_{t-n+1},\\dots,I_{t-1},I_t\\}\\).","section":"III-A"},{"comment":"The third contribution bullet ends with \"dynamic real-world environments upon paper acceptance,\" which is missing a preposition and is not a complete sentence; please rewrite.","section":"I"},{"comment":"The sentence \"with the best checkpoint (all reached before the final epoch) selected for testing\" is unclear; please state the exact checkpoint selection criterion.","section":"III-D"},{"comment":"The parameter counts are formatted inconsistently (\"6,006 k\" vs \"6,006\"), and \"GFLOPS\" should be defined or written as GFLOPs; please also clarify whether the FLOPs count includes the full multi-frame input stack.","section":"Table VII"},{"comment":"The conclusion says \"up to 8% relative improvement,\" but Table IV reports 0.873 vs 0.797 on MOT20Det (about 9.5% relative) and Table IX reports much larger absolute gains on BOAT360; please correct the percentage to match the reported numbers.","section":"V"},{"comment":"The paper promises to release the BOAT360 dataset \"upon paper acceptance\" but gives no URL or data availability statement; please provide one and, if possible, make the data available to reviewers.","section":"IV-A"}],"recommendation":"major_revision","confidential_remarks":"The empirical claim is plausible, but the current evidence is insufficient for acceptance at a serious journal. The authors should be asked to add external video-detection baselines, multi-seed statistics, and a validation-based protocol for selecting the temporal window. The BOAT360 dataset release should be available to reviewers. No concerns about conflicts of interest or citation behavior beyond the absence of a direct comparison to the video detection methods cited in Section II-B."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: a clean empirical study of a simple idea—stack n frames at the input, supervise only the latest frame—and it shows real gains for lightweight YOLOv7-tiny, especially on the new BOAT360 dataset. The paper is honest about its scope: no theoretical novelty, just a cheap practical win. That claim mostly holds.\n\nWhat is actually new: the single-frame-supervision protocol applied to multi-frame stacks for YOLO detectors, and the BOAT360 fisheye boat benchmark. The early-fusion versus grouped-convolution ablation is useful, and the parameter and speed numbers (6.011M params, 55 FPS on Orin) support the practicality argument. Internal results are consistent in direction: lightweight models gain more than heavy ones.\n\nSoft spots are real but do not sink the paper. The headline effect sizes come from post-hoc selection of the temporal window. On MOT20Det, 5-frame input is worse than 3-frame or 7-frame, and the best n is picked after seeing test results; on BOAT360, the best n is 3, not 7. No error bars or multi-seed runs are reported, and BOAT360 has only 509 total boxes, so the +12.8 mAP for YOLOv7-tiny could well be high-variance. There are also no comparisons to any existing video object detection method—the paper cites YOLOV but does not benchmark against it—and code/data are not yet released. These gaps leave the size of the effect, not its existence, in question.\n\nWho this is for: applied vision researchers working on edge deployment in robotics, surveillance, or small-boat perception. It deserves a serious referee: the idea is simple enough to be reproducible, the dataset is a useful contribution, and the missing baselines and statistics are fixable in revision. I would ask the authors for a validation-based model selection, multiple seeds, a comparison to at least one video detector, and release of the BOAT360 annotations. If those land, this becomes a solid conference paper.\n\nRecommendation: let it into peer review. It is not a desk reject.","headline":"Simple multi-frame early fusion with single-frame supervision gives lightweight YOLO detectors a real but loosely quantified boost; the effect sizes need error bars and baselines before publication.","tokens_in":9586,"tokens_out":2173,"would_cite":false,"duration_ms":24004,"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":"Stacking several consecutive video frames as input to a YOLO detector, with supervision only on the newest frame, improves detection robustness for lightweight models while adding almost no parameters.","keywords":["multi-frame detection","temporal context","YOLOv7","early fusion","lightweight models","video object detection","fisheye dataset","weak supervision"],"falsifier":"Train YOLOv7-tiny on stacks of three frames whose order is randomly permuted or taken from unrelated videos; if mAP@0.5 on MOT20Det matches the consecutive-stack result, then temporal ordering is not the source of the gain and the central claim fails.","tokens_in":1524,"feed_emoji":"🎥","tokens_out":1726,"duration_ms":96792,"temperature":0.7,"pith_summary":"The paper asks whether temporal context can make single-frame object detectors more robust without adding complex video modules. Its answer is yes: stack several consecutive frames along the channel dimension, feed them through the same YOLO backbone, and supervise only the output for the latest frame. On crowded pedestrian scenes and on a new fisheye boat dataset, this simple change raises mean average precision substantially for lightweight YOLOv7-tiny, narrowing the gap to the heavier YOLOv7. The cost is tiny, with parameter growth below 0.1 percent and inference speed that remains real-time. The paper also contributes BOAT360, an annotated fisheye video dataset captured from a moving boat, as a benchmark for this setting.","feed_headline":"Stacking video frames lifts compact YOLO detectors 12.8 mAP points","feed_subtitle":"Feeding a few consecutive frames and labeling only the newest improves detection under blur and occlusion at real-time speed.","key_machinery":"The load-bearing mechanism is input-level temporal stacking via channel concatenation: n frames become a $(3n, H, W)$ input with consistent geometric augmentation across the stack, and the first convolution's weights are initialized by repeating pretrained single-frame weights n times and rescaling by $1/n$. This lets the network fuse motion cues at the first layer, keeps parameter growth negligible, and makes sparse supervision natural because only the latest frame has labels. The paper contrasts this early fusion with a grouped-convolution variant that processes each frame separately before merging; early fusion wins consistently, showing that temporal mixing should happen at the pixel level.","core_discovery":"The central claim is that a detector does not need a temporal module to exploit video; it only needs the raw pixels of neighboring frames as extra input channels. By concatenating n consecutive RGB frames into a $(3n, H, W)$ tensor and modifying the first convolutional layer, a standard YOLOv7 or YOLOv7-tiny network can learn motion and appearance cues directly from pixels. Only the most recent frame in the stack is supervised, so unlabeled frames act as free context. Experiments on MOT20Det and BOAT360 show gains concentrated in lightweight models: YOLOv7-tiny rises from 79.7 to 87.3 mAP@0.5 with sparse 3-frame sampling on MOT20Det and from 41.1 to 53.9 mAP@0.5 with 3 adjacent frames on BOAT360, while the larger YOLOv7 gains less. The paper concludes that early pixel-level fusion of a few frames, rather than late or grouped fusion, is what carries the improvement.","pith_inferences":["Beyond the paper: the same early-fusion recipe should transfer to other single-frame backbones, since it changes only the input channels and the supervision mask; this is the authors' stated future direction, not a result shown here.","Beyond the paper: an adaptive frame-count selector would be a natural next step, because the optimal stack size shifts with scene dynamics, and a fixed stack will always be wrong for some regime.","Beyond the paper: a shuffled-frame control, training on stacks of unrelated or temporally scrambled frames, would isolate whether the gains come from temporal ordering or simply from seeing more augmented pixel data; the paper does not run this control.","Beyond the paper: BOAT360's fisheye moving-camera setting could serve as a stress test for video object detection methods beyond YOLO, since its rapid scene changes expose the limits of long temporal windows."],"forward_implications":["Lightweight detectors can close much of the accuracy gap to heavy detectors by using temporal context: on MOT20Det, multi-frame YOLOv7-tiny approaches single-frame YOLOv7 while staying much smaller.","The number of stacked frames must be tuned per deployment: on MOT20Det 7 adjacent frames peak, on BOAT360 3 frames peak, and 9 frames consistently hurt, so the temporal window is a hyperparameter rather than a universal choice.","Sparse stepped sampling matches or beats dense stacking: 3 frames with step 3 span as much time as 7 adjacent frames with fewer inputs and reach the best mAP@0.5 on MOT20Det.","Fusion should happen early: grouped convolutions that keep each frame's early features separate consistently underperform standard early fusion.","The overhead is negligible, with parameter counts growing by less than 0.1 percent and inference staying at 53-55 FPS on an embedded GPU, and the paper releases the annotated BOAT360 fisheye dataset to support further work."],"supporting_citations":[{"why":"Supplies the YOLOv7 and YOLOv7-tiny base architectures whose first convolutional layer the method adapts.","marker":"[1]"},{"why":"Provides the MOT20Det benchmark used for the central pedestrian detection experiments.","marker":"[7]"},{"why":"Defines the COCO-style mAP metrics used to measure detection performance.","marker":"[26]"},{"why":"Introduces grouped convolutions, the baseline fusion variant the paper compares against early fusion.","marker":"[25]"},{"why":"Represents the flow-guided video detection approach whose complexity the simple stacking strategy is designed to avoid.","marker":"[3]"},{"why":"Documents earlier early-fusion attempts limited to two frames or heavy 3D CNN models, motivating the lightweight extension.","marker":"[14]"}],"fun_headline_variants":["Stack 3 frames, boost tiny YOLO by 12.8 mAP","No temporal modules, just stacked frames for better YOLO","Pixel-level frame fusion lifts small YOLO detectors","3-frame stacking boosts YOLO-tiny by 12.8 mAP","Feed neighbors to YOLO: robust video detection without extra models"],"cache_read_input_tokens":11904,"weakest_assumption_plain":"That feeding a fixed set of recent frames, all shifted and scaled the same way, gives the detector useful information about the target frame; the paper's own results show this works only for certain stack sizes and scene speeds.","fun_headline_variants_meta":{"raw":{"variants":["Stack 3 frames, boost tiny YOLO by 12.8 mAP","No temporal modules, just stacked frames for better YOLO","Pixel-level frame fusion lifts small YOLO detectors","3-frame stacking boosts YOLO-tiny by 12.8 mAP","Feed neighbors to YOLO: robust video detection without extra models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000625,"raw_usage":{"total_tokens":2897,"prompt_tokens":952,"completion_tokens":1945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1851}},"tokens_in":568,"tokens_out":1945,"duration_ms":15134,"temperature":1.0,"reasoning_tokens":1851,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:45:31.074365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train YOLOv7-tiny on stacks of three frames whose order is randomly permuted or taken from unrelated videos; if mAP@0.5 on MOT20Det matches the consecutive-stack result, then temporal ordering is not the source of the gain and the central claim fails.","supporting_citations":[{"cited_title":"Flow-guided feature aggregation for video object detection,","cited_arxiv_id":null,"evidence_quote":"Represents the flow-guided video detection approach whose complexity the simple stacking strategy is designed to avoid."},{"cited_title":"Video visual relation detection via 3d convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Documents earlier early-fusion attempts limited to two frames or heavy 3D CNN models, motivating the lightweight extension."}],"review_version":1}