{"id":"2bc79070-06ed-4b6e-af4a-e8afb4a8aeb6","arxiv_id":"2411.17251","paper_version":8,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A YOLO11 plus dynamic graph neural network pipeline is claimed to detect and track small occluded traffic objects with precision 0.8382 and mAP@0.5:0.95 0.6476, but the supporting evidence is inconsistent and lacks tracking evaluation.","lead":"This paper proposes DGNN-YOLO, a traffic surveillance system that pairs YOLO11 object detection with a dynamic graph neural network for tracking small, occluded objects such as pedestrians and motorbikes. The authors report strong detection numbers on a 50,000-image traffic dataset, but the tracking claims are not measured with tracking metrics and several reported scores contradict each other.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported metrics are mutually inconsistent: Table 2/abstract (mAP@0.5:0.95=0.6476) contradict §4.4.5 (0.575) and the ablation table (mAP@0.5=0.716 vs 0.7830), so the headline outperformance cannot be verified without released checkpoints and a single reproducible evaluation pass.","rationale":"I read the paper in good faith: it aims to show that DGNN-YOLO outperforms detection baselines and uses a DGNN to track small occluded objects. For that central claim to hold, the reported final metrics must be stable and reproducible. They are not. Section 4.4.5 explicitly describes final mAP@0.5:0.95 as 0.575, while the abstract and Table 2 report 0.6476, a 12.6-point discrepancy. The ablation table gives the full framework mAP@0.5 = 0.716, precision = 0.776, and recall = 0.636, differing from Table 2's 0.7830/0.8382/0.6875. Section 4.4.3 adds yet another value, 0.812 average mAP@0.5. No held-out test set is used, and the same validation split appears to drive both training decisions and final reporting. Thus the paper's strongest claim, significant outperformance, cannot be assessed from the manuscript as written. I agree with the reader's REJECT verdict, but I weight the metric inconsistency as more foundational than the DGNN training-loop concern: if the numbers are not stable, the claimed advantage is unverifiable even before examining how the DGNN is trained. The concrete check, re-evaluating a released checkpoint with a fixed protocol, would settle whether one set of numbers is genuine; without releasing artifacts, the paper remains unverified. Credit is due for stating the XAI equations and loss formulations, but that does not compensate for the contradictory results and missing reproducibility artifacts.","tokens_in":21590,"tokens_out":5346,"duration_ms":47001,"concrete_test":"Ask the authors to release the final trained checkpoint and the exact evaluation script, then run a single evaluation pass over the same 10,000-image validation split with fixed IoU and confidence settings and compute precision, recall, mAP@0.5, and mAP@0.5:0.95. If the output does not match Table 2 (0.8382/0.6875/0.7830/0.6476) and cannot also explain Section 4.4.5 (0.716/0.575) and Table 3 (0.716/0.776/0.636) as different checkpoints or thresholds, the paper's central performance claim is contradicted by its own results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim is not internally consistent. Table 2 reports the final DGNN-YOLO as precision 0.8382, recall 0.6875, mAP@0.5 0.7830, and mAP@0.5:0.95 0.6476, and the abstract repeats these values. But Section 4.4.5 ('mAP Improvements over Epochs') states that the model reaches mAP@0.5 = 0.716 and mAP@0.5:0.95 = 0.575 by the final epoch. The ablation study in Table 3 gives the full framework mAP = 0.716, precision = 0.776, and recall = 0.636. A third set appears in Section 4.4.3, where the average mAP@0.5 is reported as 0.812. These numbers cannot all describe the same trained model on the same validation split. Moreover, the dataset section (4.1) defines only an 80/20 train/validation split with no held-out test set, and Section 4.2 says training hyperparameters were 'dynamically adjusted,' implying model selection was performed on the very split used to report final numbers. Because the paper's headline claim of 'significantly outperforming existing methods' rests entirely on these reported metrics, the inconsistency is load-bearing: the reader cannot determine which result is real, and the claimed detection advantage is unverifiable. This is independent of the additional, valid concern about the undefined DGNN training loop in Section 3.6 and Algorithm 1, which already makes the tracking claim unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGNN-YOLO, a pipeline that combines YOLO11 object detection with a dynamic graph neural network for detecting and tracking small, occluded traffic objects, and adds Grad-CAM, Grad-CAM++, and Eigen-CAM for interpretability. The central claim is that DGNN-YOLO achieves precision 0.8382, recall 0.6875, and mAP@0.5:0.95 0.6476, \"significantly outperforming existing methods,\" while running in real time. The experiments use the i2 Object Detection Dataset with an 80/20 train/validation split, and report detection metrics, an ablation study, FPS, error metrics, and qualitative XAI comparisons.","tokens_in":47,"tokens_out":4136,"duration_ms":69441,"significance":"If the claims were supported, the paper would offer a practically relevant integration of dynamic graph reasoning with a real-time detector, plus an interpretability layer, for a realistic traffic-surveillance setting. The paper also contains useful elements: an explicit algorithm sketch, a defined combined loss, a quantified ablation, and a comparison of three XAI methods. However, the central performance claim is not verifiable from the manuscript as written. The reported numbers are mutually inconsistent, no held-out test set is used, tracking is never evaluated with multi-object-tracking metrics, and the DGNN training loop is not specified. These are load-bearing issues because the paper's headline contributions—significant outperformance and robust tracking—rest on exactly this evidence. The manuscript in its current form does not provide a sound basis for accepting those claims.","major_comments":[{"comment":"The central numerical claim is internally inconsistent. Table 2 reports the final DGNN-YOLO as precision 0.8382, recall 0.6875, mAP@0.5 0.7830, and mAP@0.5:0.95 0.6476, and the abstract repeats these values. Section 4.4.5, however, states that the model reaches mAP@0.5 = 0.716 and mAP@0.5:0.95 = 0.575 by the final epoch. The ablation study in Table 3 gives the full framework mAP = 0.716, precision = 0.776, and recall = 0.636. A third set appears in Section 4.4.3, where the average mAP@0.5 is reported as 0.812. These numbers cannot all describe the same trained model on the same validation split. Because the headline claim of significant outperformance rests entirely on these metrics, the inconsistency is load-bearing: the reader cannot determine which result is real, and the claimed detection advantage is unverifiable without a single reproducible evaluation pass or released checkpoints.","section":"Section 4.4.1 (Table 2), Section 4.4.3, Section 4.4.5, Section 4.4.4 (Table 3)"},{"comment":"The experimental protocol does not support the reported generalization claim. Section 4.1 defines only an 80/20 train/validation split with no held-out test set, and Section 4.2 says training hyperparameters such as learning rate and batch size were \"dynamically adjusted\" during the experiment. This implies model selection was performed on the very split used to report the final numbers. With no separate test set and no fixed protocol, the reported precision/recall/mAP values cannot be interpreted as an unbiased estimate of performance on new data, and comparisons to baselines are confounded by whatever selection procedure was used.","section":"Section 4.1 and Section 4.2"},{"comment":"The tracking component is not specified as a trainable system. Section 3.6 defines a tracking loss L_track, but the paper never explains how the dynamic graph is constructed during training, how L_track is backpropagated through the threshold-based edge insertion/removal in Algorithm 1 (lines 18–23), how tracking IDs are assigned and maintained, or how the \"DGNN module trained from scratch\" (Section 4.2) receives gradients from YOLO11 detections. Without this training loop, the claimed tracking improvement is unsupported. The paper must specify the full forward/backward path or clarify that the DGNN is used only as a post-processing association heuristic.","section":"Section 3.6 and Algorithm 1"},{"comment":"Tracking is never evaluated with tracking metrics. The evaluation section defines precision, recall, mAP, and interpretability metrics, and the results tables report detection metrics, FPS, and error metrics. There is no MOTA, IDF1, HOTA, or identity-switch count. Since the abstract and introduction explicitly claim \"adaptive and robust tracking\" and the framework's contribution is presented as detection plus tracking, the absence of any standard multi-object-tracking evaluation makes the tracking claim untestable.","section":"Section 4.3 and Section 4.4"},{"comment":"Baseline training details are missing. Table 2 compares Faster R-CNN, YOLO5, YOLO8, YOLO9, YOLO10, and YOLO11, but the paper does not state the training epochs, learning rates, batch sizes, input resolutions, data augmentation, or any other protocol for these baselines. Without matching training conditions, the reported margin of DGNN-YOLO over YOLO11 (for example, mAP@0.5 rising from 0.6107 to 0.7830) cannot be attributed to the proposed DGNN component rather than to different training choices.","section":"Section 4.4.1, Table 2"}],"minor_comments":[{"comment":"The model name is used inconsistently as YOLO11, YOLOv11, and YOLO11-DGNN; the same notation should be used throughout.","section":"Throughout"},{"comment":"The tracking loss uses f^{t+1}_i and f^t_i, but these feature vectors are not defined in Section 3.1.3; the notation should be introduced explicitly.","section":"Equation (2)"},{"comment":"The related work refers to \"the approaches developed by Zhang et al.\" without a citation, and the description of the proposed graph pruning algorithm does not appear in the methodology.","section":"Section 2.3"},{"comment":"The sentence about YOLOv3 and YOLOv4 has an empty citation; the reference list should be completed.","section":"Section 2.1"},{"comment":"The DGNN-YOLO training time of 118,344 seconds is roughly seven times larger than that of YOLO11 (17,040 seconds), but the paper still claims real-time operation based only on inference FPS; the distinction between training cost and inference speed should be stated explicitly.","section":"Section 4.4.1, Table 2"},{"comment":"Table 4 uses a configuration \"DGNN w/o Spatial,\" but Table 3's ablation removes appearance embeddings, velocity similarity, temporal features, or edge-weight adaptivity; the relationship between these ablations should be clarified.","section":"Section 4.4.4 and Table 4"}],"recommendation":"reject","confidential_remarks":"The paper claims a strong empirical result but the reported numbers are mutually inconsistent, the evaluation uses no held-out test set, the tracking claim is not measured with MOT metrics, and the DGNN training loop is underspecified. These are not presentation issues; they undermine the central claims as written. Even with a major revision, the authors would need to rerun the evaluation under a clean protocol and make the metrics consistent, which goes beyond local corrections. I also note that no code or checkpoints are provided, which further limits verification. I recommend rejection, while acknowledging that the underlying idea of coupling YOLO with a dynamic graph for tracking is not without interest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, bottom line: the reader's skepticism is warranted. The assembled system—YOLO11 plus threshold-based dynamic graph plus off-the-shelf XAI—is a legitimate engineering exercise, and the i2 dataset gives it a concrete testbed. Applying a standard graph update rule with distance/velocity edge heuristics to traffic surveillance is not a conceptual leap, but it is a sensible thing to try, and the ablation study in Table 3 tries to isolate the contributions of appearance, velocity, and temporal features. That is more than many such papers do.\n\nThe soft spots, however, are load-bearing. The headline numbers in the abstract and Table 2 (mAP@0.5 0.7830, mAP@0.5:0.95 0.6476) conflict with Section 4.4.5, which gives final-epoch values of 0.716 and 0.575, and with Table 3's full-framework mAP of 0.716. Section 4.4.3 adds a third number (average mAP@0.5 0.812). These cannot all describe the same model on the same data. On top of that, there is no test split—just an 80/20 train/validation split—and training hyperparameters were \"dynamically adjusted,\" so model selection appears to have happened on the validation set that produces the reported metrics. Tracking is never evaluated with MOT metrics; the only tracking loss L_track is described in Equation 2, but the paper never explains how gradients flow through the thresholded graph construction during training or how tracking IDs are assigned. The claim of robust tracking is therefore not supported by the evidence presented.\n\nThis is a paper that could become a useful empirical baseline if the authors fixed the reporting: one reproducible evaluation pass, a real test set, MOT metrics, and released weights/configuration. As it stands, I would not spend referee time on it; the inconsistent metrics alone justify a desk reject. The appropriate outcome is reject, and I'd tell the authors that the path to acceptance is to rebuild the evaluation, not to patch the prose.","headline":"A reasonable engineering combo, but the reported numbers don't agree with each other, so the headline result is unverifiable as written.","tokens_in":22480,"tokens_out":2474,"would_cite":false,"duration_ms":22053,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DGNN-YOLO claims that coupling YOLO11 with a dynamically updated graph of detections raises small, occluded traffic-object detection and tracking above current detectors while keeping the system real-time and interpretable.","keywords":["dynamic graph neural network","small object detection","object tracking","occlusion","YOLO11","explainable AI","traffic surveillance","interpretability"],"falsifier":"Run the same DGNN-YOLO pipeline on MOTChallenge or KITTI with ground-truth track IDs and compute MOTA, IDF1, and HOTA against a simple baseline of YOLO11 detections matched by a nearest-neighbor tracker. If the graph module does not reduce identity switches relative to that baseline, or if removing the DGNN from the training loss leaves tracking metrics unchanged, the claimed tracking improvement is unsupported. A more direct check is to inspect gradient flow from $L_{\\text{track}}$ back to graph construction, which the paper leaves unspecified.","tokens_in":21362,"feed_emoji":"🚦","tokens_out":5164,"duration_ms":47112,"temperature":0.7,"pith_summary":"This paper argues that adding a dynamic graph neural network to YOLO11 makes small, occluded traffic objects easier to detect and track. The framework, DGNN-YOLO, treats each detected object as a node and updates edges from proximity, velocity, and appearance similarity, allowing the graph to follow objects as they move and occlude one another. On the i2 traffic dataset, the authors report precision of 0.8382, recall of 0.6875, and mAP@0.5:0.95 of 0.6476, above the YOLO11 baseline and other detectors, at roughly 60 frames per second. If this holds, a single camera-only pipeline could improve surveillance and intelligent-transportation monitoring in dense urban scenes. The paper also claims Grad-CAM, Grad-CAM++, and Eigen-CAM make the model's decisions visible and therefore more trustworthy in practice.","feed_headline":"Graph-augmented YOLO11 lifts small-object precision to 0.8382","feed_subtitle":"A dynamic graph of detections lifts recall to 0.6875 and mAP@0.5:0.95 to 0.6476 on the i2 benchmark.","key_machinery":"The load-bearing mechanism is the dynamic graph construction and update loop. Each detected object becomes a node with feature vector $x_i=[F_s,F_t]$, and edges are inserted or removed by thresholds on distance, velocity difference, and appearance similarity; the adjacency matrix $A_t$ is recomputed every frame so the graph follows objects entering, leaving, occluding, and moving. A multi-layer graph convolution propagates information through neighboring nodes, refining embeddings that the paper says are used to assign tracking IDs. The second support is the combined loss, which explicitly ties detection to temporal identity consistency and is inspired by detection-plus-re-identification tracking losses.","core_discovery":"The central claim is that detection and tracking improve when they are coupled through a continuously updated graph rather than treated as separate frame-by-frame stages. DGNN-YOLO builds one graph per frame: nodes are YOLO11 detections carrying spatial, motion, and appearance features, and edges encode Euclidean proximity, velocity similarity, and cosine appearance similarity. At each frame the graph adds or removes nodes and edges, then a graph convolution layer $H^{(l+1)}=\\sigma(A_t H^{(l)} W^{(l)})$ refines node embeddings and assigns stable tracking IDs. Training minimizes a combined loss $L=\\lambda_{\\text{det}}L_{\\text{det}}+\\lambda_{\\text{track}}L_{\\text{track}}$, where the tracking term penalizes inconsistency between connected nodes and between the same object across consecutive frames. The reported improvements are measured with detection metrics, and the paper presents the same pipeline as delivering tracking plus interpretability in real time.","pith_inferences":["Because the reported evaluation uses detection metrics (precision, recall, mAP) rather than tracking metrics such as MOTA, IDF1, or HOTA, the claimed tracking improvement is not yet directly measured; a fair test would require ground-truth track IDs and identity-switch counts.","A clean way to isolate the graph module's contribution would be to hold YOLO11 fixed and vary only edge thresholds or edge features; if the precision and mAP shifts match the ablation trends, the graph module is responsible, and if not, the gains may come from the detector re-training.","The reported training time for the full model is about seven times that of YOLO11 alone, so practical deployment would need to verify whether the gains survive pruning or quantization, which the authors list as future work.","The conclusion concedes that visual-only input fails in heavy rain and fog; this implies the claimed robustness is bounded to clear-weather urban scenes, and sensor fusion would be the natural extension."],"forward_implications":["If the claims hold, detection and tracking can be handled by one graph-based pipeline instead of separate stages, since the per-frame graph supplies both detections and identity associations.","Recall is the largest reported gain over plain YOLO11 (0.6875 versus 0.5248), so the main benefit would come from not missing small occluded objects rather than from avoiding false positives alone.","The ablation results indicate that temporal features are the most critical component, so the framework's advantage depends on using frame-to-frame motion rather than appearance alone.","At roughly 60 fps in the authors' setup, the framework stays within real-time operation for surveillance cameras on the hardware tested.","The XAI heatmaps would allow operators to see which image regions drive each detection, supporting human review of automated traffic monitoring."],"supporting_citations":[{"why":"Supplies the YOLO11 detector that provides bounding boxes, features, and initial detections used to construct graph nodes.","marker":"Alif, 2024"},{"why":"Motivates dynamically updating graph structures over time for video object detection, the basis of the tracking module.","marker":"Yuan et al., 2017"},{"why":"Provides the dynamic spatial-temporal GNN approach that the paper adopts for modeling interactions across frames.","marker":"Shao et al., 2022"},{"why":"Named as the inspiration for combining detection and re-identification losses in a unified multi-object tracking objective.","marker":"Zhang et al., 2021"},{"why":"Supplies the GNN-based end-to-end data association approach that this tracking module extends.","marker":"Jiang et al., 2019"},{"why":"Provides GNN3DMOT, a prior GNN tracking method the paper compares against conceptually and claims to improve.","marker":"Weng et al., 2020"},{"why":"Offers a spatio-temporal convolutional baseline for small object detection in video that the framework aims to surpass.","marker":"Bosquet et al., 2021"},{"why":"Provides Faster R-CNN, one of the compared detection baselines in the benchmark table.","marker":"Ren et al., 2017"}],"fun_headline_variants":["Dynamic graph nets make YOLO11 see small occluded objects","YOLO11 plus dynamic graphs lifts small-object precision to 0.8382","Interpretable DGNN-YOLO tracks tiny occluded objects in real time","Graph-powered YOLO11 achieves 0.8382 precision on small objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the DGNN module is genuinely trained to refine identity associations across frames, but the paper never specifies how the dynamic graph is built during training, how the tracking loss is backpropagated through threshold-based graph construction, how tracking IDs are assigned, or how the DGNN, trained from scratch, receives gradients from YOLO11 detections.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic graph nets make YOLO11 see small occluded objects","YOLO11 plus dynamic graphs lifts small-object precision to 0.8382","Interpretable DGNN-YOLO tracks tiny occluded objects in real time","Graph-powered YOLO11 achieves 0.8382 precision on small objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1619,"prompt_tokens":1038,"completion_tokens":581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":654,"tokens_out":581,"duration_ms":5142,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:20:10.796646+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same DGNN-YOLO pipeline on MOTChallenge or KITTI with ground-truth track IDs and compute MOTA, IDF1, and HOTA against a simple baseline of YOLO11 detections matched by a nearest-neighbor tracker. If the graph module does not reduce identity switches relative to that baseline, or if removing the DGNN from the training loss leaves tracking metrics unchanged, the claimed tracking improvement is unsupported. A more direct check is to inspect gradient flow from $L_{\\text{track}}$ back to graph construction, which the paper leaves unspecified.","supporting_citations":[{"cited_title":"Temporal dynamic graph lstm for action-driven video object detection","cited_arxiv_id":null,"evidence_quote":"Motivates dynamically updating graph structures over time for video object detection, the basis of the tracking module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Offers a spatio-temporal convolutional baseline for small object detection in video that the framework aims to surpass."}],"review_version":1}