{"id":"d48431e2-cac0-44c7-84bd-b6c09ab82fd1","arxiv_id":"1908.07085","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A PointNet-style regressor predicts 2D BEV bounding boxes from LiDAR points using a double-angle orientation head and center-relative-to-mean, outperforming L-shape fitting on KITTI in reported metrics.","lead":"BoxNet trains a simple PointNet-style network to estimate a 2D bounding box directly from a bird's-eye-view LiDAR point cluster. It reports large accuracy gains over classic L-shape fitting on KITTI, though the evaluation leaves open questions about data splitting and reproducibility.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported numbers may be inflated by an object-level train/test split: KITTI frames are not separated, so the same physical object can appear in both training and test.","rationale":"I agree with the reader's conditional verdict, but I would prioritize the train/test independence issue over the ground-truth-box input issue. The method is plausible and the ablations give some internal support: the double-angle encoding is a standard fix for angle wrap-around, the relative-center prediction is reasonable, and the EdgeConv and network-size ablations are informative. The central quantitative claim, however, rests on the KITTI evaluation. The paper does not explain how the 20,000 car samples were split. If the split is at object level rather than frame or track level, the test set can contain near-duplicates of training objects (the same physical vehicle appearing in nearby KITTI frames). This would make the reported 0.14 m center error and 1.81° orientation error unrepresentative and would break the comparison with SLF, which has no access to training data. The ground-truth-box input limitation is important for real-world deployment, but it is a scope issue that the paper partly acknowledges by framing the method as an intermediate step. The split issue is a direct threat to the validity of the numbers as reported. I am not asserting that leakage definitely occurred, only that the absence of any split description, combined with no released code, means the current evidence does not support the strong quantitative claim. The concrete frame-level split test would settle this: if the improvement over SLF largely survives, the central claim stands; if it collapses, the paper needs a substantially revised evaluation before it can be accepted.","tokens_in":9770,"tokens_out":9334,"duration_ms":100448,"concrete_test":"Re-run the whole BoxNet pipeline with a strict frame-level split of the KITTI object dataset: assign all objects from each LiDAR frame (or, better, each raw-data track) to either train or test, e.g., 80/20 by frame ID, keeping the same 30-point filter and N=512 resampling. Retrain from scratch with the reported hyperparameters and recompute Tables II and III. If the car errc/|errθ|/IoU degrade toward the SLF values, the current sample-level split is leaking and the central claim is unsupported. As a cheaper first check, use KITTI raw tracklets or object track IDs to count how many same-physical-object clusters appear in both the current 15,000/5,000 splits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A extracts points from every ground-truth box and then splits the 20,000-car sample set into 15,000/5,000 (Table I), but never states that the split is at frame or track level. KITTI contains multiple objects per frame and the same object across consecutive frames; with a random sample-level split, near-duplicate point patterns can appear in both training and test. Since BoxNet is a regression network trained on these exact point patterns, leakage of this kind would inflate the claimed car center error of 0.14 m, orientation error of 1.81°, and IoU of 0.879, and would exaggerate the margin over SLF. The ground-truth-box input assumption is a real scope limitation, but it is secondary: the paper defines the task as an intermediate module. The missing split specification is the load-bearing issue for the paper's central quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BoxNet, a PointNet-style regression network that takes an unordered set of 2D bird's-eye-view (BEV) LiDAR points belonging to a single object and predicts a 2D bounding box, parameterized by center, width, length, and orientation. To handle the periodicity of an unoriented box, the orientation head regresses the double-angle vector [cos(2θ), sin(2θ)] and recovers θ by arctan2; to ease center regression, the network predicts the offset from the input point-cloud mean. The method is evaluated on KITTI 3D object data after extracting points inside ground-truth boxes, and is compared with three variants of Search-based L-shape Fitting (SLF), reporting large improvements for cars, more modest improvements for cyclists, and mixed center-error results for pedestrians. Ablations study the angle representation, the relative-center parametrization, feature concatenation, EdgeConv layers, and network width.","tokens_in":9971,"tokens_out":6932,"duration_ms":73970,"significance":"If the reported numbers are trustworthy, the paper gives useful evidence that a simple permutation-invariant network can outperform optimization-based L-shape fitting on ground-truth-box BEV point sets, especially when only partial object geometry is observed, and that the double-angle orientation target and mean-relative center parametrization are effective design choices. The paper is deliberately scoped as an intermediate module for downstream tracking or detection, not as a full detector, and the comparison against a published non-learning baseline on a public benchmark is appropriate. The ablation study is informative and the execution-time reporting is a useful addition. However, the central quantitative claim of \"significant improvement\" is not currently fully supported: the train/test split is not specified at frame or track level, no uncertainty estimates are provided, and the evaluation protocol is restricted to points extracted from ground-truth boxes.","major_comments":[{"comment":"The paper must specify how the 15,000/5,000 train/test split was constructed. KITTI contains multiple objects per frame and the same object across consecutive frames, so a random object-level split can place near-identical point patterns in both training and testing. As written, the protocol is consistent with such a sample-level split, which would inflate BoxNet's reported car center error of 0.14 m, orientation error of 1.81 degrees, and IoU of 0.879, and would exaggerate the margin over SLF. The split must be performed at frame or track level, and the manuscript should state this explicitly; if the original experiments used an object-level split, they need to be rerun with a non-overlapping split.","section":"Section IV-A, Table I"},{"comment":"All quantitative results are reported as single point estimates without standard deviations, confidence intervals, or the number of random seeds/restarts. Given that the central claim is that BoxNet significantly improves over SLF, the paper should report variability across at least several training runs and evaluated splits, especially for the smaller cyclist and pedestrian sets where stochastic effects are likely to be larger.","section":"Section IV-D, Tables II-VII"},{"comment":"The evaluation extracts points only from the ground-truth bounding boxes of KITTI annotations. Because the paper motivates BoxNet as an intermediate step for subsequent modules such as tracking, the practical claim that it \"can be applied to various scenarios\" is not validated for point clusters produced by an upstream detector or segmenter, which may contain background points, partial objects, or merged objects. The authors should either test on such imperfect clusters or clearly state this as a limitation of the current study.","section":"Section IV-A"},{"comment":"The resampling of each point cloud to N = 512 points is not described. The manuscript should state whether points are randomly subsampled without replacement, sampled with replacement when the cloud is smaller than 512, or resampled deterministically, and whether the same resampling is applied to the point clouds used by SLF. If SLF runs on the original point clouds while BoxNet runs on resampled ones, or vice versa, the comparison is not fair; if SLF runs on resampled clouds, the effect of resampling on the baseline's performance should be discussed.","section":"Section IV-A"}],"minor_comments":[{"comment":"Reference [19] lists the first KITTI author as \"A. Geger\"; the correct spelling is \"A. Geiger\".","section":"References"},{"comment":"The notation for the center error is inconsistent: Equation (2) defines it as errc, but the text before Table IV refers to errd and Table V uses errd. Please unify the notation.","section":"Section IV-C, Tables IV-V"},{"comment":"The histograms would benefit from explicit bin widths, axis labels, and a legend in the caption or figure itself; the current text does not indicate what the vertical axis represents or how the bins were chosen.","section":"Figure 3"},{"comment":"The word \"shrinked\" should be \"shrunk\" in the text and in Table VII.","section":"Section IV-F and Table VII"}],"recommendation":"major_revision","confidential_remarks":"This is a short empirical paper whose central claim is plausible but not yet fully supported because of the unspecified split level and missing uncertainty estimates. If the authors can confirm a frame- or track-level split and add error bars, I would view it as suitable for publication in a robotics or vision venue. If the split was in fact object-level, the quantitative claims would require substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a modest, honestly-written paper about a narrow intermediate step — fitting a 2D bounding box to the BEV point cloud of a single object. The architecture is a simple PointNet-style regressor, and the two tricks are predicting sin(2θ), cos(2θ) instead of θ, and predicting the box center relative to the point-cloud mean. Both are sensible and the ablations support them. The empirical comparison against SLF, a non-learning L-shape fitter, is the paper's main contribution, and the reported gains for cars are large: center error 0.14 m vs 0.39 m, orientation error 1.8° vs 8.4°, IoU 0.879 vs 0.683.\n\nThe paper does several things well. The task is cleanly defined and the evaluation metrics are standard. The ablations are informative, especially the EdgeConv experiment showing that local neighbor features do not help and the network-size study showing that a much smaller network retains most of the benefit. The authors also correctly limit orientation to (−π/2, π/2] and explain why heading is not estimable from sparse BEV points. That is honest scoping.\n\nThe soft spots are real but not fatal. The most important is the train/test split. Section IV-A extracts points from every ground-truth box and then says 15,000/5,000 car samples, but never states that the split is at frame or track level. KITTI contains multiple objects per frame and the same object across consecutive frames. If the split is random over objects, near-duplicates leak across the boundary, and that would inflate the reported margins. This is a load-bearing detail for the paper's central numbers, and it should have been specified. I do not think it invalidates the method — even with some leakage, a learned regressor should beat an L-shape fitter on partially observed objects — but the magnitude of the improvement is uncertain.\n\nSecond, the input is always the ground-truth box's points. The authors acknowledge this defines the task as an intermediate module, but it means the results do not directly transfer to clusters from an upstream detector. That is a scope limitation, not a hidden flaw. Third, no error bars, no code, and the resampling to 512 points is described only briefly. For a paper whose contribution is an empirical comparison, these are missing.\n\nThe citation pattern is fine; the related work on L-shape fitting is covered and the comparison to SLF is fair. I see no circularity problem — this is straightforward supervised regression on labeled boxes.\n\nWho is this for? Anyone working on LiDAR-based tracking or detection pipelines that need fast, accurate box refinement from object clusters. It is a useful datapoint, not a breakthrough. I would send it to review if the authors can clarify the split, add variance estimates, and release code; without those, the numbers are hard to check.\n\nMy verdict: deserves a serious referee, with the expectation of at least a major revision on the experimental reporting. If the split turns out to be frame-level, the paper is probably acceptable after a light revision.","headline":"A plausible but under-specified empirical study of learned 2D box geometry from object-level BEV points; the split ambiguity must be fixed before the headline numbers can be trusted.","tokens_in":10450,"tokens_out":2372,"would_cite":false,"duration_ms":21339,"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":"BoxNet shows that a simple point-order-invariant network can regress a 2D bounding box from bird's-eye-view LiDAR points, cutting car center error to 0.14 m and orientation error to 1.8 degrees on KITTI.","keywords":["2D bounding box estimation","bird's-eye view point cloud","LiDAR perception","deep learning","orientation regression","double-angle encoding","L-shape fitting","KITTI benchmark"],"falsifier":"Reproduce the car experiments with clusters produced by an off-the-shelf detector or segmenter on KITTI instead of ground-truth boxes; if average IoU falls toward the L-shape baseline or center error exceeds roughly 0.2 m, the reported gains depend on clean ground-truth cropping.","tokens_in":9604,"feed_emoji":"🚗","tokens_out":9478,"duration_ms":85483,"temperature":0.7,"pith_summary":"BoxNet claims that a simple deep network can replace classical L-shape fitting when estimating a 2D bounding box from bird's-eye-view LiDAR points. Given only the unordered 2D coordinates of points inside an object, the network regresses the box center, width, length, and orientation, using training data to learn typical object sizes. On the KITTI car category, this cuts average center error from about 0.39 m to 0.14 m and average orientation error from about 8.4 to 1.8 degrees, while intersection-over-union rises from 0.68 to 0.88. If the result holds, downstream modules such as tracking and data association could get faster and more accurate box geometry without hand-designed fitting rules.","feed_headline":"One small network beats L-shape box fitting at LiDAR pose estimation","feed_subtitle":"On KITTI cars, BoxNet cuts center error from 39 cm to 14 cm and angle error from 8.4 to 1.8 degrees.","key_machinery":"The load-bearing objects are the double-angle orientation encoding and the mean-relative center regression. Because heading is not estimated, the orientation lies on $(-\\pi/2, \\pi/2]$; the network regresses $(\\cos 2\\theta, \\sin 2\\theta)$, which is continuous at the wrap-around, and recovers $\\theta = \\operatorname{atan2}(\\sin 2\\theta, \\cos 2\\theta)/2$. The center head estimates $(\\tilde c_x, \\tilde c_y)$, the box center relative to the point-cloud mean, and the absolute center is obtained by adding the mean back; concatenating the predicted angle and size with the pooled feature vector improves the center estimate. The feature extractor uses shared MLPs and a max-pooling layer, so the prediction does not depend on point ordering, and the size head uses ReLU to output positive width and length learned from training data.","core_discovery":"The central claim is that the geometry of an object — center, width, length, and orientation — can be recovered from a set of unordered 2D bird's-eye-view LiDAR points by a small regression network with three heads. The orientation head outputs $(\\cos 2\\theta, \\sin 2\\theta)$ rather than $\\theta$ itself, because the bounding-box orientation is defined on $(-\\pi/2, \\pi/2]$ and the doubled-angle representation is continuous across the discontinuity at $\\pm \\pi/2$. The center head predicts the offset from the point-cloud mean, and the recovered absolute center plus learned width and length produce the box. For cars, the average center error is 0.1401 m, the orientation error is 1.8057 degrees, and the IoU is 0.8787, versus 0.3941 m, 8.4223 degrees, and 0.6825 for the best L-shape fitting baseline.","pith_inferences":["Not tested in the paper: running BoxNet on point clusters produced by a real detector or segmenter, rather than clusters cut from ground-truth boxes; because training never sees imperfect clusters, this test would show how much of the reported gain survives upstream noise.","The double-angle encoding likely transfers to any orientation-regression problem with 180-degree symmetry, such as 2D vehicle boxes in images, since it removes angle wrap-around ambiguity without needing a classification head.","The ablation with a nearest-neighbor convolution suggests boundary points carry more regression signal than local neighbor structure; weighting or explicitly sampling boundary points could improve accuracy on sparse and heavily occluded objects.","Extending the architecture to full 3D boxes by adding height while keeping the double-angle yaw encoding is a natural next step, and the same mean-relative center trick should apply in three dimensions."],"forward_implications":["For cars, average IoU rises from 0.6825 for the best L-shape fitting variant to 0.8787 for BoxNet on the same KITTI test samples, so a trained regressor aligns boxes with true extents far better than optimization-based fitting.","Because the size head learns a width/length prior from training data, BoxNet can output the full object size even when only a partial set of points is visible, a situation L-shape fitting cannot handle because it fits only the observed points.","Point-order invariance removes the need to sort, cluster, or segment points before fitting, which simplifies use in tracking and data-association pipelines.","Shrinking the network to 1/16 of its baseline size costs less than 2% average IoU, and inference takes roughly 7.4–8.5 ms on a P100 GPU, so the accuracy gain does not require a heavy model.","The improvement is class-dependent: orientation error drops sharply for all classes, but pedestrian center error stays near 0.10 m and remains comparable to SLF, because near-square boxes give no clear primary axis."],"supporting_citations":[{"why":"It supplies the shared-MLP plus max-pooling network structure that makes BoxNet point-order invariant.","marker":"[7]"},{"why":"It is the state-of-the-art non-learning L-shape fitting baseline that BoxNet is compared against.","marker":"[16]"},{"why":"It provides the deep-network 3D bounding-box regression approach from images that BoxNet adapts to 2D BEV LiDAR points.","marker":"[18]"},{"why":"It provides the KITTI 3D object dataset with annotated LiDAR frames used for training and testing.","marker":"[19]"},{"why":"It replaces early MLP layers with a nearest-neighbor convolution in the ablation study, showing that local neighbor information does not help bounding-box regression.","marker":"[22]"},{"why":"It defines the orientation range (-pi/2, pi/2] that BoxNet adopts by ignoring object heading.","marker":"[17]"}],"fun_headline_variants":["BoxNet: deep learning for LiDAR boxes from unordered points","Double-angle regression improves LiDAR box orientation from BEV","BoxNet beats L-shape fitting on BEV LiDAR: 64% lower center error","Neural net estimates 2D LiDAR boxes from unordered points","BoxNet reduces LiDAR box errors: center 39 to 14 cm, angle 8.4 to 1.8 deg"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes every test input is a clean point cluster clipped from a ground-truth bounding box, with no upstream detector noise; if real clusters contain background, partial, or merged objects, the reported improvements over L-shape fitting may shrink.","fun_headline_variants_meta":{"raw":{"variants":["BoxNet: deep learning for LiDAR boxes from unordered points","Double-angle regression improves LiDAR box orientation from BEV","BoxNet beats L-shape fitting on BEV LiDAR: 64% lower center error","Neural net estimates 2D LiDAR boxes from unordered points","BoxNet reduces LiDAR box errors: center 39 to 14 cm, angle 8.4 to 1.8 deg"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001695,"raw_usage":{"total_tokens":6716,"prompt_tokens":948,"completion_tokens":5768,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":5660}},"tokens_in":564,"tokens_out":5768,"duration_ms":40301,"temperature":1.0,"reasoning_tokens":5660,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:26:15.542043+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the car experiments with clusters produced by an off-the-shelf detector or segmenter on KITTI instead of ground-truth boxes; if average IoU falls toward the L-shape baseline or center error exceeds roughly 0.2 m, the reported gains depend on clean ground-truth cropping.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the shared-MLP plus max-pooling network structure that makes BoxNet point-order invariant."},{"cited_title":"Zhang, W","cited_arxiv_id":null,"evidence_quote":"It is the state-of-the-art non-learning L-shape fitting baseline that BoxNet is compared against."},{"cited_title":"3D Bounding Box Estimation Using Deep Learning and Geometry","cited_arxiv_id":"1612.00496","evidence_quote":"It provides the deep-network 3D bounding-box regression approach from images that BoxNet adapts to 2D BEV LiDAR points."},{"cited_title":"Geger, P","cited_arxiv_id":null,"evidence_quote":"It provides the KITTI 3D object dataset with annotated LiDAR frames used for training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the orientation range (-pi/2, pi/2] that BoxNet adopts by ignoring object heading."}],"review_version":1}