{"id":"e5098ac2-6fcc-4a2f-a674-8bbc5d11f18c","arxiv_id":"2504.18203","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A monocular camera pipeline with LiDAR-guided depth training detects railway objects in 3D up to 250 meters, but the final end-to-end 3D accuracy is not quantitatively reported.","lead":"This paper builds a train-mounted camera system that detects obstacles in 3D out to 250 meters using only a single camera at run time, while using LiDAR recordings to train the depth part of the model. For railway automation, it is one step toward the much longer perception range trains need.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No end-to-end 3D metric supports the 250-meter claim; the full-pipeline evaluation reports only 2.5D mAP and depth MAE, with large test-time distance errors (e.g., buffer stop MAE 119.75 m) that would corrupt frustums at long range.","rationale":"The reader's REJECT verdict is correct. My pass isolates the same decisive weakness: the paper claims long-range monocular 3D detection but never measures it end-to-end on the target dataset. The reader frames this via the depth-accuracy assumption, citing the authors' own statement that frustum quality is degraded by 2.5D and depth errors. I agree with that mechanism, but I would put more weight on the more direct evidentiary gap: there is no end-to-end 3D mAP or recall figure on OSDaR23 at all. The reported 3D numbers come from component evaluations that use ground-truth 2D detections and LiDAR/inpainted depth inputs, so they cannot validate the monocular test-time pipeline. The large test-time distance errors in Table VI make it implausible that the pipeline delivers accurate 3D boxes at 250 m, but the stronger point is that this is unmeasured. The train/test overlap admitted in Section IV-B further weakens even the module-level evidence. Because my concern reinforces rather than displaces the reader's rationale, the verdict remains REJECT, i.e., no change from the reader's verdict. I selected 'partial' rather than 'agree' because the reader's weakest-assumption wording emphasizes depth error as an assumed risk, whereas I see the load-bearing issue as the complete absence of any full-pipeline 3D metric, which is a direct failure of evidence rather than an untested assumption.","tokens_in":11059,"tokens_out":3547,"duration_ms":35218,"concrete_test":"Assemble the MFF pipeline exactly as described (monocular image -> 2.5D YOLOv9 detections + estimated depth -> pseudo-cloud -> frustumization and decision -> short/long-range heads), run it on the OSDaR23 test frames from sequences disjoint from training, and compute per-class BEV and 3D mAP at IoU 0.1 (and 0.25 if applicable) over all ranges and over the 100-150 m, 150-200 m, and 200-250 m bins, using the same frustum-based evaluation protocol as Faraway-Frustum. Report recall separately for the long-range head. If no such end-to-end 3D numbers can be produced, or if 3D mAP at 200-250 m is near zero, the abstract's central claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's statement that the monocular pipeline detects objects up to 250 meters on OSDaR23. To support this, the assembled pipeline must be shown to produce usable 3D boxes at that range. Section IV-B, however, reports no end-to-end 3D detection metric. The OSDaR23 test-set evaluation contains only 2.5D mAP/MAE (Table VI) and depth MAE (12.85 m), plus qualitative images. The only 3D metrics in the paper (Tables III, IV, V) are for isolated components: the long-range head is trained and tested with ground-truth 2D labels and OSDaR23 point-cloud/depth inputs, not with the monocular pipeline's predicted 2.5D detections and estimated depth. The authors explicitly acknowledge in Section IV-B that 'the error carried from the 2.5D predictions and depth maps can end up affecting the quality of the frustums, which at the same time ends up hindering 3D object predictions.' The reported errors are not small in the regime of the claim: the 2.5D module's buffer stop MAE on the test set is 119.75 m, and the overall MAE is 15.76 m; even a 12.85 m depth error at 200-250 m substantially shifts pseudo-cloud geometry. Also, Section IV-B states there is 'significant overlap between training and test sequences,' meaning even the module-level numbers are likely optimistic. Thus the most load-bearing gap is not a hypothesized failure but an unmeasured one: the full monocular pipeline has never been evaluated quantitatively for 3D detection, so the 250-meter claim is not established by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a modular monocular 3D object detection pipeline for railway scenes, termed MFF (Monocular Faraway-Frustum). The pipeline combines a YOLOv9-based 2.5D object detector, a DenseDepth/Marigold-based monocular depth estimation module, frustum generation with a weighted centroid/distance decision rule, and separate short- and long-range 3D detection heads (evaluated with PointRCNN, PointPillars, and Part-A2 for short range; an FF-Net-style head for long range). LiDAR point clouds are used only during training to supervise depth refinement and to generate frustums for head training. The authors report module-level validation results on OSDaR23 and a KITTI comparison for the long-range head, and the abstract claims that the method detects objects up to 250 meters. The full-pipeline evaluation on the OSDaR23 test set, however, reports only 2.5D detection metrics (Table VI) and depth MAE, with 3D detections shown only qualitatively in Figure 6.","tokens_in":11360,"tokens_out":3265,"duration_ms":33340,"significance":"If the headline claim were quantitatively supported, the work would be a valuable contribution to long-range perception for autonomous rail, a domain with comparatively few public datasets and methods. The paper has genuine strengths: it uses a relevant real-world dataset (OSDaR23), provides component-level ablation-style comparisons, includes a KITTI benchmark for the long-range head, and openly discusses limitations such as depth-map noise and label imbalance. The modular design is clearly described and reproducible in principle. However, the central claim that the full monocular pipeline detects objects up to 250 meters is never measured end-to-end in 3D. The only quantitative full-pipeline results are 2.5D mAP/MAE and depth MAE, and the authors acknowledge both substantial per-class distance errors (buffer stop MAE 119.75 m) and significant overlap between training and test sequences. These issues directly undermine the abstract's central claim and the practical relevance of the approach.","major_comments":[{"comment":"The abstract claims that evaluations on OSDaR23 demonstrate detecting objects up to 250 meters, but the full-pipeline evaluation contains no end-to-end 3D detection metric. Table VI reports only 2.5D mAP@0.5 and MAE, and Figure 6 is qualitative. Consequently, the central claim is not quantified for the assembled monocular pipeline at any range, let alone at 200-250 m.","section":"Section IV-B, Table VI and Figure 6"},{"comment":"The long-range head is evaluated using ground-truth 2D labels and OSDaR23-derived depth maps (LiDAR/inpainted), not the monocular pipeline's predicted 2.5D boxes and estimated depth. Table III therefore measures the head in isolation under idealized inputs. The KITTI comparison in Table IV uses a different domain and class (Car) and does not validate the railway long-range setting. These numbers cannot support the claim that the assembled pipeline detects objects at 250 m.","section":"Section III-D / Section IV-A.3, Table III"},{"comment":"The authors state that performance on the test set closely matches validation 'given the significant overlap between training and test sequences.' This is a data-leakage concern: if test frames come from the same sequences as training frames, the reported test-set results (Table VI, depth MAE 12.85 m) are optimistic and cannot be interpreted as generalization performance. The paper does not quantify the overlap or provide results on disjoint sequences.","section":"Section IV-B, paragraph on test-set overlap"},{"comment":"The authors acknowledge that 'the error carried from the 2.5D predictions and depth maps can end up affecting the quality of the frustums, which at the same time ends up hindering 3D object predictions.' Given the empirically reported errors (e.g., buffer stop MAE 119.75 m at test time, full MAE 15.76 m, depth MAE 12.85 m), the effect on long-range frustum geometry is likely severe. This admission, combined with the absence of an end-to-end 3D metric, means the 'up to 250 meters' claim is not merely unproven but is contradicted by the reported error magnitudes.","section":"Section IV-B, discussion of error propagation"}],"minor_comments":[{"comment":"The phrase 'an deep-learning-based approach' should be 'a deep-learning-based approach'.","section":"Abstract / Introduction"},{"comment":"The table caption and text would benefit from explicitly stating that the long-range head evaluation uses ground-truth 2D labels and OSDaR23 depth maps, so the reader does not mistake these for end-to-end results.","section":"Section IV-A.3, Table III"},{"comment":"The assessment states that PointPillars was chosen for the full pipeline, but no end-to-end comparison of the three short-range heads in the assembled pipeline is reported; a sentence explaining why the integration-level comparison was omitted would clarify the evaluation strategy.","section":"Section IV-A.4, Table V"},{"comment":"Several minor typographical errors and inconsistencies appear (e.g., 'utlized', 'Associationation', 'V oixel', 'GTs' vs 'GT'); a careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper would be more suitable as a technical report or a workshop presentation than as a journal article in its current form. The central claim is unsupported by the reported experiments, and the test-set overlap raises validity concerns. A resubmission with a proper end-to-end 3D evaluation on disjoint sequences, including per-range breakdowns, would be needed to assess the approach's actual long-range capability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a system-integration paper for monocular 3D object detection in railway settings, built on the Faraway-Frustum idea but replacing LiDAR at inference with pseudo-clouds from a learned depth map. The individual pieces – a YOLOv9 with a distance head, DenseDepth distilled from Marigold, and short/long-range 3D heads – are known, but the specific combination for OSDaR23, with LiDAR used only in training, is new. Credit where due: component-level evaluation is reasonably careful, the long-range head comparison against Faraway-Frustum on KITTI (60.61 vs 46.90 3D mAP@0.1) is a legitimate check, and the authors are transparent about error accumulation between modules.\n\nThe main soft spot – exactly as the stress-test note says – is that the headline claim of detecting objects up to 250 meters is never quantified for the full pipeline. Section IV-B reports only 2.5D mAP/MAE and depth MAE on the test set, plus qualitative images. The only 3D numbers come from isolated heads fed with ground-truth 2D labels and dataset depth/point clouds, not from the monocular pipeline's own outputs. A buffer stop MAE of 119.75 m at test time makes it hard to believe long-range frustums survive intact. The authors themselves acknowledge this: 'the error carried from the 2.5D predictions and depth maps can end up affecting the quality of the frustums.' They also admit 'significant overlap between training and test sequences,' which means even the module-level numbers are likely optimistic. No code, no error bars, single runs. There is also a nagging mismatch between the stated motivation – trains need perception beyond 1 km – and a system that stops at 250 m, though that is a scope issue rather than a flaw in what they did.\n\nThat said, I would not desk reject this. The problem is real, the design is sensible, and the paper reads like an honest workshop submission whose abstract overreaches. A serious referee could ask for an end-to-end 3D metric on a clean train/test split, or a revised claim that says 'the pipeline shows promising 2.5D detection at range, and components achieve 3D detection when given clean inputs.' I would send it to review with that expectation. The stress-test note holds up.\n\nFor you: if you work on railway perception or long-range monocular 3D detection, it is worth a skim for the module comparisons and OSDaR23 details. I would not cite it for the 250-meter capability until that capability is actually measured end-to-end.","headline":"Sensible modular pipeline for monocular railway 3D detection, but the 250-meter claim is not backed by any end-to-end 3D metric; the authors are honest about the gap, making this a workshop-grade paper that needs either a full-pipeline 3D evaluation or a scaled-back claim.","tokens_in":11979,"tokens_out":2764,"would_cite":false,"duration_ms":27748,"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":"A monocular camera, trained with LiDAR depth guidance, can detect and localize 3D railway objects up to 250 meters away.","keywords":["monocular 3D object detection","railway perception","depth estimation","frustum-based detection","LiDAR-guided training","long-range detection","OSDaR23","autonomous trains"],"falsifier":"Measure per-object depth error and 3D detection recall on the OSDaR23 test set separately for the 200-250 meter range: if the full pipeline's 3D recall at that range drops to near zero when predicted depth maps are used instead of ground-truth frustums, the 250-meter detection claim would be disproven.","tokens_in":10818,"feed_emoji":"🚆","tokens_out":3875,"duration_ms":36000,"temperature":0.7,"pith_summary":"This paper claims that a single camera, trained with LiDAR depth supervision, can perform 3D object detection in railway scenes at distances up to 250 meters. The proposed pipeline, MFF, replaces the LiDAR point cloud used by the Faraway-Frustum approach with pseudo-clouds generated from learned monocular depth, so that no LiDAR is needed at test time. Evaluations on the OSDaR23 dataset show detection of persons, road vehicles, buffer stops, catenary poles, and signal poles at long range, and the long-range head outperforms Faraway-Frustum's head on KITTI car detection (60.61 vs. 46.90 3D mAP@0.1). This matters because trains require braking distances far exceeding those of cars, so extending perception beyond the typical automotive range is safety-critical.","feed_headline":"Camera-only rail detector reaches 250 meters with LiDAR-guided training","feed_subtitle":"LiDAR supervises depth only during training; at test time a single camera yields 3D boxes up to 250 meters.","key_machinery":"The load-bearing mechanism is the pseudo-cloud frustum: monocular depth maps are back-projected into 3D pseudo point clouds, and 2.5D object detections carve out pyramid-shaped frustums that are routed to dedicated short- and long-range detection heads. This replaces the LiDAR point cloud that the original Faraway-Frustum needed while preserving the frustum's role of narrowing the 3D search space. The long-range head consumes bird's-eye-view splatted frustums with class priors, and the short-range head consumes raw frustum points via a LiDAR-based detector, with PointPillars selected as the best performer.","core_discovery":"The central claim is that LiDAR guidance during training alone can enable a monocular-only pipeline to produce frustum-based 3D detections at long range. MFF distills relative depth from a diffusion-based teacher (Marigold) into a real-time DenseDepth variant, then fine-tunes a refinement network to absolute depth using inpainted LiDAR ground truth. A modified YOLOv9 with an added distance head outputs 2.5D detections, and a weighted sum of frustum centroid and estimated distance routes each frustum to a short-range LiDAR-based head or a long-range BEV head. On the OSDaR23 test set the full pipeline detects objects up to 250 meters, and the long-range head reaches 60.61 3D mAP@0.1 on KITTI cars versus 46.90 for Faraway-Frustum. The authors state that depth-map and 2.5D errors propagate into frustum quality and currently limit 3D prediction accuracy.","pith_inferences":["The 250-meter ceiling may be a property of depth quality more than of the frustum architecture; reducing depth error at range could push the same pipeline toward the kilometer-scale perception trains ultimately need.","The scarcity of long-range training samples in OSDaR23 suggests that performance beyond 200 meters could improve substantially with more balanced annotations or synthetic augmentation.","The weighted-sum routing between short- and long-range heads could be replaced by a learned, end-to-end mechanism, since the paper itself notes that module errors propagate through frustumization.","A testable extension would apply the same LiDAR-guided depth distillation to other camera-based 3D detectors to separate the contribution of depth quality from that of the frustum design."],"forward_implications":["If correct, camera-only systems could provide early 3D warnings at 250 meters without requiring per-vehicle LiDAR at deployment, lowering cost for railway automation.","The long-range head architecture transfers to automotive settings, as shown by its improvement over Faraway-Frustum's head on KITTI.","Distilling depth from a diffusion teacher into a real-time network offers a path to accurate absolute depth estimation without ground-truth depth labels.","The modular pipeline implies that improving the 2.5D distance head or the depth estimation module would directly improve downstream 3D detection quality.","Demonstrating long-range detection on OSDaR23 provides a benchmark for future railway perception research beyond the automotive domain."],"supporting_citations":[{"why":"Faraway-Frustum is the multi-modal framework this paper adapts, providing the dual-head design and BEV long-range head.","marker":"[21]"},{"why":"OSDaR23 is the railway dataset used for training and evaluation, supplying images, LiDAR, and long-range 3D labels.","marker":"[46]"},{"why":"YOLOv9 is the base detector modified with a distance head to create the 2.5D detection module.","marker":"[26]"},{"why":"Marigold is the diffusion-based teacher network used to distill relative depth into the real-time student.","marker":"[35]"},{"why":"DenseDepth provides the encoder-decoder architecture that is modified and distilled for monocular depth estimation.","marker":"[32]"},{"why":"KITTI supplies pre-training data for the 2.5D detector and the validation benchmark for the long-range head comparison.","marker":"[24]"},{"why":"Dist-YOLO serves as the comparison baseline for 2.5D detection and distance estimation performance.","marker":"[29]"},{"why":"The colorization-using-optimization method is used to inpaint sparse LiDAR depth maps into dense ground-truth depth for supervisor training.","marker":"[34]"}],"fun_headline_variants":["LiDAR-trained camera sees 3D objects at 250m","Monocular 3D detection reaches 250m via LiDAR guidance","Camera-only rail detection hits 250m with LiDAR teacher","LiDAR-guided depth training enables 3D boxes at 250m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that depth maps predicted from a single image, especially beyond 200 meters, are accurate enough to build frustums that preserve the 3D structure of objects; the paper itself acknowledges that errors from 2.5D predictions and depth maps degrade frustum quality and hinder 3D predictions.","fun_headline_variants_meta":{"raw":{"variants":["LiDAR-trained camera sees 3D objects at 250m","Monocular 3D detection reaches 250m via LiDAR guidance","Camera-only rail detection hits 250m with LiDAR teacher","LiDAR-guided depth training enables 3D boxes at 250m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2694,"prompt_tokens":945,"completion_tokens":1749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1673}},"tokens_in":561,"tokens_out":1749,"duration_ms":11612,"temperature":1.0,"reasoning_tokens":1673,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:21:56.472978+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure per-object depth error and 3D detection recall on the OSDaR23 test set separately for the 200-250 meter range: if the full pipeline's 3D recall at that range drops to near zero when predicted depth maps are used instead of ground-truth frustums, the 250-meter detection claim would be disproven.","supporting_citations":[{"cited_title":"Faraway-Frustum: Dealing with Lidar Sparsity for 3D Object Detec- tion using Fusion,","cited_arxiv_id":null,"evidence_quote":"Faraway-Frustum is the multi-modal framework this paper adapts, providing the dual-head design and BEV long-range head."},{"cited_title":"Open sensor data for rail 2023,","cited_arxiv_id":null,"evidence_quote":"OSDaR23 is the railway dataset used for training and evaluation, supplying images, LiDAR, and long-range 3D labels."},{"cited_title":"Dist-yolo: Fast object detection with distance estimation,","cited_arxiv_id":null,"evidence_quote":"Dist-YOLO serves as the comparison baseline for 2.5D detection and distance estimation performance."},{"cited_title":"Colorization using optimiza- tion,","cited_arxiv_id":null,"evidence_quote":"The colorization-using-optimization method is used to inpaint sparse LiDAR depth maps into dense ground-truth depth for supervisor training."}],"review_version":1}