{"id":"8a4281a2-3cf6-4f9b-b883-705eba66b246","arxiv_id":"2412.18165","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"This paper describes a two-network parallel GPU architecture for LiDAR scene understanding in racing, claiming a 2x speedup, but the perception evaluation is circular and unvalidated.","lead":"This paper proposes running two independent neural networks on separate GPUs to process LiDAR data for autonomous racing, and reports a 2x inference speedup over a sequential setup. A specialist might read it as a systems baseline, but the perception results are not validated against any external ground truth.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Circular benchmark: untrained segmentation output is used as ground truth (Section IV-B), so the reported reconstruction accuracies and scene-understanding claims have no external validity; only the expected 2x hardware speedup remains.","rationale":"The reader's weakest-assumption diagnosis is correct and is directly supported by the manuscript text. The paper's scientific claims about segmentation and reconstruction depend on treating the untrained segmentation network's output as ground truth; this is a circular benchmark because the 'ground truth' is itself an arbitrary random-initialization output. The reconstruction network's accuracy numbers therefore cannot be interpreted as measuring scene understanding, and the qualitative comparison between loss functions is not evidence of external validity. The 2x speedup is a separate, more plausible systems result, but it is trivially expected when doubling the number of GPUs and does not compensate for the invalidated perception claims. The paper does release code and weights, which is good and makes the proposed concrete test feasible. I found no need to adjust the reader's REJECT verdict; the central argument is not supported by the evidence as presented.","tokens_in":8614,"tokens_out":4705,"duration_ms":44124,"concrete_test":"Using the released HuggingFace weights, run the trained reconstruction network on a held-out RACECAR sequence and compute IoU of its output against the actual input BEV map (raw LiDAR occupancy), not against the segmentation network's output. Also train a control reconstruction network with the same architecture and loss using the actual BEV map as ground truth. If the pseudo-label-trained model does not beat this control on actual-BEV IoU, or if its apparent accuracy collapses when scored against true occupancy, the reported 99%/61% numbers are circular artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step for every scene-understanding claim is Section IV-B: 'Due to the lack of hand-labelled annotations in the RACECAR dataset, we employ the segmentation network's output as ground truths for the corresponding input sequence to train the reconstruction network.' The segmentation network is never trained for segmentation; Section III and Section IV-C explicitly state that it segments 'without any training required.' A randomly initialized encoder-decoder with skip connections is not a segmentation function; its output is an arbitrary function of the input and random weights. Therefore the reconstruction network is trained to reproduce a pseudo-label, not to reconstruct the scene. Table I's 99%/61% reconstruction accuracies and the qualitative comparison between IoU loss and MSSCE loss measure agreement with that pseudo-label, so they cannot support any claim about scene understanding. The segmentation network's own 98.2% figure is also undefined (no metric, no data split, no repetitions) and is evaluated on a different task, predicting future BEV maps, not segmenting the current scene. Removing these unsupported perception claims leaves Table II's 2x latency reduction, which is the expected consequence of running two independent networks on two separate GPUs rather than sequentially on one. That systems result may be valid, but it does not establish the paper's central claim about parallel neural computing for scene understanding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Parallel Perception Network (PPN) for autonomous racing, consisting of two independent encoder-decoder convolutional networks — a 'segmentation' network with skip connections and a 'reconstruction' network without them — that run concurrently on two separate NVIDIA T4 GPUs. The input is a sequence of 2D Bird's Eye View (BEV) maps converted from LiDAR point clouds. The authors claim a 2x inference-time speedup relative to a sequential configuration, and report reconstruction accuracies of 99% and 61% under IoU and MSSCE losses, respectively, plus a 98.2% accuracy for the segmentation network. The paper also introduces the MSSCE loss, which combines MSE, SmoothL1, and Canny-edge-preserving terms. Training and evaluation are carried out on a single RACECAR scenario (PoliMove Multi-Agent Slow on LVMS), with 7,150 LiDAR sweeps.","tokens_in":8816,"tokens_out":3205,"duration_ms":30809,"significance":"If the system-level result held, the paper would offer a simple, reproducible demonstration that two independent perception networks can be placed on separate GPUs to reduce inference latency — a modest but useful contribution for latency-sensitive autonomous racing. The authors also make code and learned parameters publicly available, which is commendable and improves reproducibility of the systems part. However, the scene-understanding claims are not supported by the evaluation: the reconstruction network is trained and tested against the output of an untrained, randomly initialized segmentation network, making the reported accuracies measures of agreement with a pseudo-label rather than with any ground-truth scene structure. The segmentation network's own accuracy is undefined (no metric, split, or repetitions), and the only quantitative evidence that survives this criticism is the latency comparison, which is itself underreported and partially inconsistent with the 'at least two times' claim in Section IV-D. The paper's central title claim about scene understanding therefore rests on a circular benchmark.","major_comments":[{"comment":"The evaluation of the reconstruction network is circular. Section IV-B states: 'Due to the lack of hand-labelled annotations in the RACECAR dataset, we employ the segmentation network's output as ground truths for the corresponding input sequence to train the reconstruction network.' However, the segmentation network is never trained for segmentation; Sections III and IV-C explicitly state that it segments 'without any training required.' A randomly initialized encoder-decoder with skip connections produces an arbitrary function of the input and random weights, not a meaningful scene representation. Therefore, Table I's 99% and 61% reconstruction accuracies merely quantify how well the reconstruction network reproduces this pseudo-target, and the qualitative comparison between IoU and MSSCE losses (Fig. 7) cannot support any claim about scene reconstruction or scene understanding. Since the input BEV maps themselves could have served as ground truth for the autoencoder-style reconstruction task, this is not a fundamental data-annotation limitation but a methodological choice that invalidates the reported perception metrics.","section":"Section IV-B and Table I"},{"comment":"The segmentation network's reported accuracy of 98.2% is undefined and unsupported. No metric (e.g., IoU, pixel accuracy) is specified, no data split or number of repetitions is reported, and no comparison against manually annotated or otherwise validated ground truth is provided. Furthermore, the benchmark described in Section IV-B — training a modified version of the segmentation network to predict future BEV maps — is a different task from segmenting the current scene; the 98.2% figure cannot be the accuracy of the untrained segmentation network on that task, because Table I's row for 'Segmentation Net' is listed after training. The text also does not explain how the untrained network could yield any quantitative accuracy without a defined target. This row of Table I should be removed or replaced with a clearly specified evaluation protocol.","section":"Section IV-C and Table I"},{"comment":"The latency measurement is underreported and the 'at least two times' speedup claim is contradicted by the table's own numbers. Section IV-D says the parallel configuration 'reveal[s] a speedup of at least two times,' but Table II lists sequential min/max of 0.162/0.205 seconds and parallel min/max of 0.075/0.091 seconds. The worst-case speedup is 0.162/0.091 ≈ 1.78×, which is below 2×. Additionally, the measurement protocol is not described: no number of runs, no variance, no warm-up, no explanation of whether BEV conversion and data transfer are included, and no specification of whether the sequential configuration ran both networks on one GPU or on two GPUs time-shared. Without this information, the 2× claim is plausible but not established. The paper should report mean, standard deviation, and the exact measurement methodology, and the 'at least' qualification should be corrected.","section":"Table II and Section IV-D"}],"minor_comments":[{"comment":"There are several typographical errors, including 'infernece' in the Table II caption, 'comparision' in the Section IV-E heading, and 'seperate' in Table II. These should be corrected.","section":"Throughout"},{"comment":"Figure 6 is described as showing 'PPN model's input and outputs without training,' but the caption does not state which panels correspond to which network or how the blank reconstruction output is produced. Please clarify the figure legend and refer to it explicitly in the text.","section":"Figure 6"},{"comment":"The BEV conversion parameters are incompletely specified: the paper mentions a 1000×1000 grid and a voxel 'resolution/voxel size' in Algorithm 1 but does not give the numerical voxel size or the LiDAR range used for cropping. Please provide these values for reproducibility.","section":"Section IV-A"},{"comment":"The definition of MSSCE as simply the sum LMSE+Canny + LSmoothL1+Canny means each pixel is counted twice, and it is unclear whether the authors intended a weighted average. If the two terms are equally weighted, this should be stated explicitly; if not, the formula is ambiguous.","section":"Equation (5)"}],"recommendation":"major_revision","confidential_remarks":"The perception evaluation is circular because the segmentation network used to generate pseudo-ground-truth is never trained, so all reconstruction accuracy numbers lose external validity. This is not a minor presentation issue; it undermines the 'scene understanding' contribution. However, the parallel-execution idea and the latency measurement, if properly reported, could form a legitimate systems note. I would be willing to consider a revised manuscript that reframes the contribution as a parallel-execution baseline — dropping or explicitly re-scoping the segmentation claims, training the reconstruction network against the input BEV maps, and providing a rigorous latency benchmark. The current version, as written, does not establish the scene-understanding claims, but the core systems idea is salvageable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2412.18165.\n\nThe paper's only concrete result is that running two independent networks on two GPUs halves inference time compared to running them sequentially on one GPU. That is true but trivially expected; it is not a contribution. The scene-understanding claims do not hold up. The authors train the reconstruction network to reproduce the segmentation network's output, and the segmentation network is never trained; it has random weights and skip connections. Section IV-B states this explicitly. So Table I's 99% and 61% accuracies measure agreement with a pseudo-label, not any external truth. The segmentation network's 98.2% figure is also undefined: it comes from a separately trained future-prediction head, not from the untrained segmentation, and no metric or data split is given.\n\nWhat is genuinely good: the paper is clear about what it does, and it releases code and pretrained weights. The MSSCE loss is a sensible combination of existing losses, though not novel. The authors also acknowledge the hardware cost of parallelism in Table III. But none of that rescues the central evaluation.\n\nThe timing measurement in Table II lacks run counts, variance, warm-up, and any statistical treatment. Given the expected result, this is a minor issue compared to the circularity.\n\nWhom is this for? A reader who wants a minimal example of multi-GPU inference with BEV maps might find the code useful. As a scientific claim about scene understanding, it should not be taken seriously. The paper would not pass peer review in its current form; the evaluation design would need external ground truth (e.g., hand-labeled segmentation or a real reconstruction error against LiDAR occupancy). I would not cite it. I would not bring it to reading group. A desk reject is appropriate.","headline":"The perception evaluation is circular because the reconstruction network's ground truth is the output of an untrained segmentation network, leaving only an expected 2x multi-GPU speedup that lacks measurement detail.","tokens_in":9377,"tokens_out":2728,"would_cite":false,"duration_ms":24228,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two independent LiDAR networks running in parallel on separate GPUs cut perception latency in half for autonomous racing.","keywords":["autonomous racing","LiDAR perception","bird's eye view","parallel neural networks","encoder-decoder","GPU acceleration","scene understanding","point cloud"],"falsifier":"Run the same two networks with identical batch sizes and input sizes on one NVIDIA T4 GPU in sequence and on two T4 GPUs in parallel, measuring wall-clock latency over many trials; if the parallel configuration does not show roughly a 2x speedup, the central performance claim is refuted.","tokens_in":8343,"feed_emoji":"⚡","tokens_out":9466,"duration_ms":77474,"temperature":0.7,"pith_summary":"This paper proposes a baseline architecture for hardware-enabled parallel neural computing in autonomous racing, where LiDAR perception must keep up with vehicles moving at up to 274 km/h. The Parallel Perception Network (PPN) splits scene understanding into two independent encoder-decoder networks—segmentation and reconstruction—that run concurrently on separate GPUs, converting raw 3D point clouds into 2D bird's-eye-view maps before inference. The authors report a 2x speedup in inference time (0.075–0.091 s vs 0.162–0.205 s) when the two networks run in parallel on two NVIDIA T4 GPUs compared with running them sequentially on one GPU. If the speedup holds, it shows that multi-task perception can be made faster without redesigning the networks, simply by exploiting available parallel hardware.","feed_headline":"Two GPUs cut LiDAR perception time in half for racing","feed_subtitle":"Running each network on its own GPU cuts inference to 0.075–0.091 s, roughly half the sequential 0.162–0.205 s.","key_machinery":"The central object is the Parallel Perception Network (PPN), a pair of encoder-decoder CNNs: a segmentation network with skip connections and a reconstruction network without them. Both take a stack of 16 consecutive binary bird's-eye-view maps generated from LiDAR point clouds by voxelization and max-pooling along the z-axis. The temporal dimension is processed by a pseudo-1D convolution with kernel size $T \\times 1 \\times 1$ on the input, and each network runs on its own NVIDIA T4 GPU; because the networks are independent and never share activations during inference, running them on separate devices halves the latency of the combined perception step. The reconstruction network is trained with a combined Mean Square Smooth Canny Edge (MSSCE) loss that blends MSE, SmoothL1, and an edge-preserving term based on Canny edge detection.","core_discovery":"The central discovery is that two independent convolutional networks—one with skip connections for segmentation, one without for reconstruction—can be trained and deployed in parallel on separate GPUs, yielding a roughly two-fold reduction in wall-clock inference time (0.075–0.091 s versus 0.162–0.205 s) with no change to the per-network architecture. The segmentation network produces space-time segmentations of a sequence of LiDAR-derived BEV maps without any training on labeled data; its output is then used as ground truth to train the reconstruction network, which lacks skip connections and must learn to reproduce the input scenes. The authors present this 'true hardware-enabled parallelism' as a scalable baseline for multi-network, multi-sensor perception in high-speed autonomous driving.","pith_inferences":["A natural extension is to benchmark the same two networks on a single GPU as one multi-task network with a shared encoder; a speedup that is purely proportional to added hardware says little about model efficiency, only about resource allocation.","Because the reconstruction network is trained against the untrained segmentation outputs, the reported 99% and 61% accuracies should be read as measures of self-consistency between the two networks, not as agreement with human-annotated scene labels; a labeled evaluation set would be needed to establish external validity.","The parallel scheme could be tested for synchronization overhead at higher input frame rates or with networks that must exchange intermediate features, where the clean 2x speedup would likely degrade.","If the same parallel design is applied to camera and LiDAR networks, the speedup may differ because of varying input sizes and preprocessing costs; measuring this would reveal whether the 2x result generalizes beyond the specific setting."],"forward_implications":["Multi-network perception models can be accelerated by assigning each independent network to a separate GPU, without altering the network designs.","A LiDAR perception pipeline that previously ran at roughly 5–6 Hz on one GPU could run at roughly 11–13 Hz in the parallel configuration, closer to the sensor's 10–30 Hz frame rate.","The architecture can be extended to additional independent networks (e.g., one per sensor modality) and still maintain per-network inference latency, provided each has dedicated hardware.","The untrained segmentation network's output can serve as training targets for a sibling reconstruction network, removing the need for hand-labeled annotations in the RACECAR dataset.","Using the edge-preserving MSSCE loss produces sharper reconstructions than IoU loss, though with a lower IoU score (61% vs 99%)."],"supporting_citations":[{"why":"Supplies the RACECAR LiDAR dataset used to train and evaluate the networks.","marker":"[9]"},{"why":"Provides the U-Net encoder-decoder with skip connections on which the segmentation network is based.","marker":"[16]"},{"why":"Introduces spatio-temporal perception from BEV maps using 2D convolutions, the approach PPN adapts.","marker":"[21]"},{"why":"Provides the feature pyramid network idea that shapes the encoder-decoder hierarchy.","marker":"[10]"},{"why":"Defines the MSCE edge-preserving loss that PPN extends to the MSSCE loss.","marker":"[15]"},{"why":"Defines the nuScenes LiDAR scan format and devkit used to load RACECAR point clouds.","marker":"[1]"},{"why":"Supplies the BEV mapping technique for converting 3D point clouds to 2D maps.","marker":"[2]"}],"fun_headline_variants":["Parallel neural nets halve LiDAR inference time for racing","Two GPUs enable parallel perception nets, doubling LiDAR speed for racing","Racing AI: two independent networks on separate GPUs cut LiDAR time in half","Parallel Perception Network delivers 2x faster LiDAR scene understanding","Split LiDAR networks across GPUs for 2x faster perception in racing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the untrained segmentation network's output is a valid ground truth for training and evaluating the reconstruction network; if that output does not reflect the actual scene structure, the reported reconstruction accuracies do not measure real-world performance.","fun_headline_variants_meta":{"raw":{"variants":["Parallel neural nets halve LiDAR inference time for racing","Two GPUs enable parallel perception nets, doubling LiDAR speed for racing","Racing AI: two independent networks on separate GPUs cut LiDAR time in half","Parallel Perception Network delivers 2x faster LiDAR scene understanding","Split LiDAR networks across GPUs for 2x faster perception in racing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1475,"prompt_tokens":940,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":556,"tokens_out":535,"duration_ms":5271,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:58:19.554940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same two networks with identical batch sizes and input sizes on one NVIDIA T4 GPU in sequence and on two T4 GPUs in parallel, measuring wall-clock latency over many trials; if the parallel configuration does not show roughly a 2x speedup, the central performance claim is refuted.","supporting_citations":[{"cited_title":"Racecar-the dataset for high-speed autonomous racing","cited_arxiv_id":null,"evidence_quote":"Supplies the RACECAR LiDAR dataset used to train and evaluate the networks."},{"cited_title":"Motionnet: Joint perception and motion prediction for autonomous driving based on bird’s eye view maps","cited_arxiv_id":null,"evidence_quote":"Introduces spatio-temporal perception from BEV maps using 2D convolutions, the approach PPN adapts."},{"cited_title":"Msce: An edge-preserving robust loss function for improving super-resolution algorithms","cited_arxiv_id":null,"evidence_quote":"Defines the MSCE edge-preserving loss that PPN extends to the MSSCE loss."},{"cited_title":"Bevmap: Map-aware bev modeling for 3d perception","cited_arxiv_id":null,"evidence_quote":"Supplies the BEV mapping technique for converting 3D point clouds to 2D maps."}],"review_version":1}