{"id":"d828ae21-ab5a-4807-9f45-095d5a9ee704","arxiv_id":"2507.11910","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SEPose provides the first synthetic event-based pedestrian pose estimation dataset for fixed traffic cameras, with 73K frames and roughly 350K pedestrian instances.","lead":"SEPose is a synthetic dataset of about 73,000 event-camera frames from CARLA, annotated with pedestrian body keypoints in traffic scenes. The paper trains pose and detection models on it and tests them on real event-camera data, with noticeable drops in accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim-to-real evidence is box-level detection, not pose estimation: RVT and YOLOX are detection models and no keypoint metric is reported, so the central pose-generalization claim is currently unsupported.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: box-level mAP from pose-derived boxes does not validate keypoint localization. My stress-test confirms this is not merely a weak assumption but a central mismatch between the paper's claim and its evidence. The models are detection architectures, the evaluation metric is detection mAP, and PEDRo lacks pose annotations entirely, so no pose-specific generalization result exists. This is a correctness risk for the abstract's strongest claim, not a stylistic issue. The dataset itself may still be useful, and the sim-to-real evidence for detection is real, so the condition should be to add keypoint-level evaluation—on DHP19 for real data and on a held-out synthetic split for internal validity. The instance-count inconsistency (350K keypoints vs. 350K pedestrians) and lack of public release are secondary; the missing pose evaluation is the gate for accepting the paper's headline. I therefore agree with the reader's CONDITIONAL verdict and do not escalate to rejection, because the dataset contribution is plausible and the missing experiments are addressable.","tokens_in":8190,"tokens_out":3371,"duration_ms":38040,"concrete_test":"Train a pose-estimation model (e.g., YOLOv8-pose or ViTPose) on SEPose keypoint annotations; evaluate it on a held-out synthetic split using COCO-style OKS/AP, then evaluate the same model on DHP19's projected 2D keypoints (using the paper's camera-projection procedure) with the same keypoint metrics. If keypoint AP on DHP19 is near zero or far below the reported box mAP, the sim-to-real pose generalization claim is refuted; if keypoint AP is substantial, the claim gains support.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that SEPose supports sim-to-real generalization for human pose estimation—is not tested by the reported experiments. Section IV states: 'we generate bounding boxes from all pose annotations and compute the mean average precision and mAP50 scores for the corresponding IoU scores.' The two evaluated models are RVT with a YOLO detection head and YOLOX, both object detectors; neither outputs keypoints. No keypoint-level metric (OKS, PCK, or keypoint AP) is reported on any split, synthetic or real. Thus the 69.9 mAP on PEDRo demonstrates pedestrian/box detection transfer, not pose estimation transfer. The mismatch is compounded because PEDRo has only bounding-box annotations (no pose keypoints), so keypoint-level evaluation on PEDRo is impossible; DHP19 provides 3D joints that can be projected to 2D, but the paper does not use them for keypoint accuracy. Consequently, the abstract's 'sim-to-real generalization capabilities' is unsupported for the pose task the dataset is named for. Even the dataset's internal validity for pose learning (e.g., held-out synthetic keypoint AP) is unmeasured, so the reader cannot tell whether models trained on SEPose learn any pose structure at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SEPose, a synthetic event-based multi-person human pose estimation dataset for fixed traffic-camera pedestrian monitoring, generated with the CARLA simulator. The dataset comprises about 73K event frames, corresponding RGB frames and raw event streams, and 16 COCO-style 2D body keypoints per annotated pedestrian, with diversity across six CARLA maps, weather conditions, time-of-day, crowd densities, and camera elevations. The authors describe the generation and cleaning pipeline, including weather parameter sweeps, collision-triggered resets, ray-cast occlusion filtering, and a minimum bounding-box area filter. They then train RVT with a YOLO detection head and YOLOX on SEPose and evaluate on the real event-based datasets PEDRo and DHP19, reporting mAP and AP50. The central claim is that training on SEPose demonstrates sim-to-real generalization for event-based human pose estimation. My assessment is that the dataset creation is described in valuable detail, but the experimental evidence does not currently support the pose-estimation generalization claim because only box-level detection metrics are reported.","tokens_in":8450,"tokens_out":2990,"duration_ms":35818,"significance":"If the pose-generalization claim were supported, SEPose would fill a clear gap: a publicly described, large-scale synthetic event-based dataset with pedestrian pose keypoints for ITS-oriented, fixed-camera monitoring, together with a detailed generation recipe that others could replicate in CARLA. The paper's strengths are its concrete pipeline description (sensor configuration, weather sweeps, occlusion filtering via ray casting, bounding-box filtering, Gaussian crowd-density sampling) and the choice of diverse maps and camera placements. The dataset contribution is potentially useful regardless of the transfer results. However, the claimed sim-to-real generalization for human pose estimation is not supported by the reported experiments, which evaluate only person/box detection and do not report any keypoint-level metric on synthetic or real data. The paper also contains a numerical inconsistency between the abstract and Section III-D concerning the dataset size, which must be resolved.","major_comments":[{"comment":"The central claim of sim-to-real generalization for human pose estimation is not supported by the reported experiments. Section IV states: 'we generate bounding boxes from all pose annotations and compute the mean average precision and mAP50 scores for the corresponding IoU scores.' The two models evaluated, RVT with a YOLO detection head and YOLOX, are object detectors that output bounding boxes, not keypoints. No keypoint-level metric (OKS, PCK, or keypoint AP) is reported on any split, and no held-out synthetic SEPose evaluation is presented. Because PEDRo only has bounding-box annotations, and DHP19's projected 2D keypoints are used only to form boxes, the reported 69.9 mAP on PEDRo and 51.1 mAP on DHP19 demonstrate detection transfer, not pose transfer. The authors must either add a proper keypoint evaluation on DHP19, add an in-domain held-out pose evaluation on SEPose, or explicitly revise the claims in the abstract and conclusion to describe detection-level transfer.","section":"Abstract and Section IV"},{"comment":"There is a factor-of-16 discrepancy in dataset size reporting. The abstract says 'nearly 350K annotated pedestrians,' while Section III-D says the dataset contains 'over 350K human pose keypoints.' Since each pedestrian is annotated with 16 keypoints, 350K keypoints corresponds to roughly 22K pedestrian instances, not 350K pedestrians. The authors should correct the abstract, Section III-D, and any other occurrences, and state clearly whether the intended quantity is the number of pedestrian instances or the total number of keypoints.","section":"Section III-D and Abstract"},{"comment":"The absence of any in-domain evaluation on SEPose itself weakens the empirical claims. The only numerical results are on PEDRo and DHP19, so the reader cannot distinguish a sim-to-real domain gap from a model's failure to learn pose structure from SEPose. At minimum, the paper should report keypoint-level performance on a held-out synthetic split (e.g., keypoint AP or PCK), which would establish that models trained on SEPose actually learn the annotated pose task before the transfer question is addressed.","section":"Section IV"}],"minor_comments":[{"comment":"The text says YOLO8X is trained on event frames generated 'with a time delta of 0.33 ms,' which appears to be a typographical error: the event accumulation time and simulation step described in Section III-A are 33 ms. Please correct the unit or value.","section":"Section IV"},{"comment":"The model naming is inconsistent: the abstract refers to YOLOv8, while Table III and the text use 'YOLO8X' and 'Yolo8X,' and Reference [34] is actually YOLOX. Please standardize the model name and ensure the cited paper matches the evaluated architecture.","section":"Abstract and Section IV"},{"comment":"The discussion says PEDRo 'includes more diverse indoor as well as outdoor scenarios,' but Section II describes PEDRo as an outdoor event-based dataset recorded with a handheld camera. These descriptions are inconsistent and should be reconciled.","section":"Section IV"},{"comment":"Reference [25] is cited for the COCO keypoint annotation format, but that reference is a pose-estimation paper, not the COCO dataset or keypoint challenge definition. Please cite the original COCO keypoint annotation source.","section":"Section III-C"},{"comment":"The paper would benefit from a short statistics table stating the total number of annotated pedestrian instances, the average number of keypoints per instance, the number of frames, and the split into maps and weather conditions; this would also resolve the quantity mismatch noted above.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The dataset generation effort is real and the pipeline description is useful, but the paper's headline claim—sim-to-real generalization for pose estimation—rests on box-level detection metrics. This is a load-bearing issue that can be fixed by adding keypoint-level evaluation on DHP19 and a held-out synthetic pose split, and by correcting the dataset-size statement. I recommend major revision rather than rejection because the resource itself appears valuable and the missing evidence is within scope to produce. The naming inconsistencies and the 0.33 ms vs 33 ms discrepancy should also be cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: SEPose is a genuinely useful new resource—the first large synthetic event-based pedestrian pose dataset for fixed traffic cameras—but the paper's evidence for sim-to-real pose generalization is currently box-level detection, not pose estimation, and the instance count in the abstract is off by about 16x.\n\nWhat's new and good: CARLA is used sensibly, with fixed traffic-camera placements at 10m and 2m, varied maps, weather sweeping, Gaussian crowd density, and a ray-cast occlusion filter that removes pedestrians with >50% keypoints occluded. The dataset format follows COCO keypoints with 16 joints, and the pipeline is described in enough detail to reproduce. Few published resources do this for event-based traffic scenes; DHP19 is indoor and PEDRo is detection-only, so the gap is real.\n\nSoft spots: the experiments don't test pose estimation. RVT with a YOLO head and YOLOX are object detectors; the paper generates bounding boxes from keypoint annotations and computes mAP. No keypoint-level metric (OKS, PCK, keypoint AP) is reported on any split, including SEPose itself. So the abstract's claim that training on SEPose demonstrates 'sim-to-real generalization capabilities' is unsupported for the pose task the dataset is named for. The conclusion that models 'learn meaningful features required for accurate predictions on real data' is an overstatement given the large accuracy drops—e.g., RVT mAP drops from 69.9 to 44.4 on PEDRo.\n\nSecond, the numbers don't line up: the abstract says 'nearly 350K annotated pedestrians,' but Section III-D says 'over 350K human pose keypoints.' With 16 keypoints per pedestrian, those differ by a factor of 16. The text should say what it actually is.\n\nThird, the dataset is not publicly accessible from the paper—no URL, no release statement. For a dataset paper, that's a real deficiency; the community can't verify or use the contribution.\n\nMinor: abstract says YOLOv8, experiments use YOLOX; should be aligned.\n\nThese are fixable. The generation pipeline is solid and the dataset itself is the contribution; the evaluation can be redone or the claims reframed as detection transfer. The paper deserves a serious referee, but it needs major revision before acceptance.\n\nRecommendation: send to peer review, but require keypoint-level evaluation on held-out SEPose and on DHP19's projected 2D keypoints, correct the counts, and make the data available.\n\nRegards.","headline":"The dataset fills a real gap, but the paper's pose-generalization evidence is box-level detection and the abstract's pedestrian count is ~16x too high.","tokens_in":8984,"tokens_out":2591,"would_cite":false,"duration_ms":27522,"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":"The paper introduces SEPose, a synthetic event-based pedestrian pose dataset, and argues that models trained on it generalize to real event-camera data well enough to support traffic-safety monitoring.","keywords":["event camera","human pose estimation","synthetic dataset","pedestrian monitoring","sim-to-real transfer","CARLA simulator","dynamic vision sensor","intelligent transportation"],"falsifier":"Recompute the same transfer experiment using keypoint-level metrics such as OKS or PCK on PEDRo and DHP19 rather than bounding boxes derived from keypoints. If the models' bounding-box mAP is high but per-keypoint localization error is close to chance, or clearly worse than a model trained on real event data, the paper's claim of meaningful pose-feature transfer is unsupported.","tokens_in":8012,"feed_emoji":"🚶","tokens_out":6628,"duration_ms":69863,"temperature":0.7,"pith_summary":"The paper introduces SEPose, a synthetic event-based human pose estimation dataset built for fixed traffic-camera pedestrian monitoring, generated in the CARLA simulator with a dynamic vision sensor. It claims that the dataset's roughly 73K event frames and 350K annotated pedestrians, spanning diverse weather, lighting, and crowd densities across urban, suburban, and rural intersections, provide enough structure for models to generalize to real event-camera data. To test this, the authors train RVT and YOLOv8 on SEPose and evaluate on the real PEDRo and DHP19 datasets. They report substantial absolute performance and argue that the transfer shows synthetic training produces meaningful pose features for real data. If correct, this gives the event-vision community a low-cost, diverse source of pedestrian pose data for safety-critical monitoring.","feed_headline":"350K synthetic pedestrians train pose models that work on real event data","feed_subtitle":"New dataset uses CARLA traffic scenes to train event-camera pedestrian pose models without real-world labeling.","key_machinery":"The central object is the dataset itself, generated by a pipeline that pairs CARLA's RGB cameras with a simulated dynamic vision sensor operating synchronously at 33ms steps. Weather parameters are incremented every 100 frames to cover rain, fog, and day-night cycles, and pedestrian spawn counts are sampled from a Gaussian with mean 50 and standard deviation 15 to vary crowd density. A ray-cast occlusion filter drops pedestrians whose keypoints are more than half hidden, and a 300-pixel bounding-box minimum removes distant figures. For evaluation, pose annotations from SEPose, PEDRo, and DHP19 are converted to bounding boxes and scored with mAP and mAP50 over IoU thresholds; this box-level protocol is the bridge between the synthetic training set and real test sets.","core_discovery":"The paper's central claim is that a large synthetic event-based dataset can supply the missing training data for pedestrian pose estimation from fixed traffic cameras. SEPose contains about 73K event frames with 350K annotated pedestrians, captured by a simulated dynamic vision sensor with 800x600 resolution and 90-degree field of view, across six CARLA maps and systematically varied weather and lighting. Annotations follow the COCO keypoint format with 16 joints, and each annotation entry is filtered to remove pedestrians with more than 50% occluded keypoints or bounding boxes smaller than 300 pixels in area. Training RVT and YOLOv8 on SEPose and evaluating on the real PEDRo and DHP19 datasets gives, respectively, mAP/AP50 of 69.9/98.1 and 63.2/94.7 on SEPose, dropping to 44.4/53.9 and 39.4/55.1 on PEDRo, and 51.1/61.3 and 53.7/58.9 on DHP19. The paper interprets these nonzero transfer numbers as evidence that models trained on synthetic data learn meaningful pose features for real event data.","pith_inferences":["A stricter test would evaluate keypoint-level agreement (for example, OKS or PCK) directly on PEDRo and DHP19, since the paper's bounding-box mAP converts poses to boxes and may not isolate keypoint localization quality.","If the sim-to-real gap is mostly due to event-camera noise and background statistics, then adding synthetic noise models or small amounts of real unlabeled event data during training could substantially close the reported drop.","Combining SEPose with the existing real datasets in mixed training could yield a model that keeps SEPose's viewpoint diversity while anchoring to real event statistics, a setup the paper does not test.","A direct real-to-synthetic evaluation (training on PEDRo/DHP19 and testing on SEPose) would clarify which aspects of the gap come from the synthetic sensor model versus the domain shift in scenes."],"forward_implications":["Event-based pose models can be pre-trained on synthetic traffic scenes, reducing the need for expensive manual annotation of real event-camera pedestrian data.","The variety of side and three-quarter pedestrian views, unlike centered indoor datasets, gives models exposure to viewpoints typical of fixed traffic cameras.","Because annotations include 16 body keypoints, downstream tasks such as gait-direction prediction and collision-risk estimation become trainable from the same data.","The measured sim-to-real gap on PEDRo and DHP19 provides a baseline for future work on event-camera noise simulation and domain adaptation."],"supporting_citations":[{"why":"Supplies the real outdoor multi-person detection dataset used as a transfer target in the sim-to-real experiments.","marker":"[7]"},{"why":"Supplies the real indoor event-based 3D human pose dataset, projected to 2D keypoints, used as the second transfer target.","marker":"[8]"},{"why":"Provides the CARLA simulator that generates the synthetic scenes, DVS events, and pedestrian behavior.","marker":"[10]"},{"why":"Defines the recurrent vision transformer architecture trained on SEPose event tensors.","marker":"[33]"},{"why":"Defines the YOLO-based architecture (Yolo8X) trained on SEPose event frames.","marker":"[34]"},{"why":"Defines the COCO keypoint annotation format used for the 16-joint ground truth in SEPose.","marker":"[25]"}],"fun_headline_variants":["Synthetic event data trains pose models for real cameras","350K simulated pedestrians enable event-camera pose transfer","CARLA events train pose models that work on real footage","Simulated event frames shrink sim-to-real pose gap","Large synthetic event set boosts pedestrian pose accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that bounding boxes generated from pose keypoints are a valid measure of pose estimation performance, so the reported sim-to-real evidence depends on box-level detection reflecting keypoint localization quality.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic event data trains pose models for real cameras","350K simulated pedestrians enable event-camera pose transfer","CARLA events train pose models that work on real footage","Simulated event frames shrink sim-to-real pose gap","Large synthetic event set boosts pedestrian pose accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1303,"prompt_tokens":953,"completion_tokens":350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":275}},"tokens_in":569,"tokens_out":350,"duration_ms":5032,"temperature":1.0,"reasoning_tokens":275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:58:50.283214+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the same transfer experiment using keypoint-level metrics such as OKS or PCK on PEDRo and DHP19 rather than bounding boxes derived from keypoints. If the models' bounding-box mAP is high but per-keypoint localization error is close to chance, or clearly worse than a model trained on real event data, the paper's claim of meaningful pose-feature transfer is unsupported.","supporting_citations":[{"cited_title":"Pedro: an event-based dataset for person detection in robotics,","cited_arxiv_id":null,"evidence_quote":"Supplies the real outdoor multi-person detection dataset used as a transfer target in the sim-to-real experiments."},{"cited_title":"Dhp19: Dynamic vision sensor 3d human pose dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the real indoor event-based 3D human pose dataset, projected to 2D keypoints, used as the second transfer target."},{"cited_title":"Carla: An open urban driving simulator,","cited_arxiv_id":null,"evidence_quote":"Provides the CARLA simulator that generates the synthetic scenes, DVS events, and pedestrian behavior."},{"cited_title":"Recurrent vision transformers for object detection with event cameras,","cited_arxiv_id":null,"evidence_quote":"Defines the recurrent vision transformer architecture trained on SEPose event tensors."},{"cited_title":"Yolox: Exceeding yolo series in 2021,","cited_arxiv_id":null,"evidence_quote":"Defines the YOLO-based architecture (Yolo8X) trained on SEPose event frames."},{"cited_title":"2d human pose estimation with explicit anatomical keypoints structure constraints,","cited_arxiv_id":null,"evidence_quote":"Defines the COCO keypoint annotation format used for the 16-joint ground truth in SEPose."}],"review_version":1}