{"id":"e1099757-83a1-488b-b0e1-285c2073d55b","arxiv_id":"1908.11789","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A lightweight two-stream CNN trained on a new fisheye surround-view dataset detects moving vehicles and pedestrians, reaching about 40% moving-object IoU versus 10% when trained on rectilinear KITTI data.","lead":"This paper introduces a neural network that detects moving objects in the 360-degree fisheye views around a car, together with a new parking-scene dataset for that task. It reports that a model trained on regular camera images only reaches 10% moving-object IoU on fisheye data, while training on fisheye images raises the score to about 40%.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All Table 1 numbers depend on unvalidated semi-automated LiDAR-derived masks; the KITTI-vs-fisheye comparison is also confounded by dataset size and scene domain.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the semi-automated annotation pipeline is trusted without validation. My stress-test confirms this is the most fundamental issue because all reported metrics are computed against those masks. I additionally note the KITTI-vs-fisheye comparison in Table 1 is confounded by training set size and scene domain, but that is secondary; even if the comparison were clean, the unvalidated ground truth would still prevent the reported accuracies from being trusted. There is no evidence of dishonesty or a fatal internal inconsistency, and the proposed architecture and dataset creation effort are plausible. The correct disposition remains conditional: the paper should be accepted only if the annotation quality is demonstrated and the comparison is made less confounded. Since the reader already reached CONDITIONAL, my read does not change the verdict.","tokens_in":5085,"tokens_out":3652,"duration_ms":33275,"concrete_test":"Manually annotate a random subset of, say, 100 test frames by multiple annotators using the original video plus projected LiDAR motion labels, then compute IoU and per-pixel precision/recall for the moving class between the manual masks and the semi-automated masks from Section 3. If moving-class mask IoU is below about 0.8, or if precision/recall is strongly unbalanced, regenerate Table 1 using corrected masks and report whether the 40% MOD IoU and the 10%-vs-40% gap survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The single most load-bearing assumption is the ground-truth quality of the Section 3 semi-automated annotation pipeline. Every IoU and mIoU in Table 1, and the paper's central claim that fisheye-specific training is necessary, is measured against binary masks generated by: parsing existing 2D object bounding boxes, selecting LiDAR points inside those boxes, classifying the point cloud as moving or static, projecting the points onto the fisheye image, and taking a convex hull. No manual verification, inter-annotator agreement, or sensitivity analysis is reported. If the LiDAR moving/static classifier is wrong, or if ego-motion compensation is imperfect, the masks will be systematically biased; because moving pixels are only 0.54% of a frame, even a small false-positive rate among static pixels can dominate the moving-object IoU. The reported mIoU of 69.5 is driven mainly by the dominant static class, so the informative number is MOD IoU = 40, and that number inherits all ground-truth errors. A secondary but real concern is that Table 1's evidence for 'rectilinear models do not generalize to fisheye' compares a model trained on 1,300 KITTI samples with one trained on 3,638 fisheye samples; dataset size and scene distribution are confounded with camera model, so the 10% vs 40% gap does not isolate fisheye geometry. The annotation issue is more fundamental because it undermines every reported accuracy value, including the headline 40% IoU.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FisheyeMODNet, a lightweight two-stream CNN for moving object detection (MOD) in surround-view fisheye images for autonomous driving. The authors collect a 5139-frame fisheye dataset from parking scenes, generate binary moving-object masks through a semi-automated LiDAR-based pipeline, and train a two-stream encoder with shared weights to output per-pixel moving/static labels. The central empirical claims are that a rectilinear-KITTI-trained MODNet generalizes poorly to fisheye images (10% MOD IoU), that training on the new fisheye dataset raises MOD IoU to 39.8-40%, that static-object augmentation improves it to 42%, and that the network runs at 15 fps on a 1 TOPS automotive embedded system.","tokens_in":5394,"tokens_out":3073,"duration_ms":28726,"significance":"If the results are reliable, the paper makes a useful contribution: it is an early attempt at MOD for 360-degree fisheye surround-view cameras, it provides a new dataset with MOD annotations, and it demonstrates a lightweight architecture with a realistic embedded-speed measurement. The design choice of feeding two sequential images directly instead of precomputed optical flow is sensible for fisheye cameras, where optical flow is distorted and requires calibration-dependent preprocessing. The weight-sharing scheme that reuses the previous encoder is a practical efficiency idea. However, the significance is conditional on the validity of the ground-truth masks and on the fairness of the rectilinear-versus-fisheye comparison; both are currently unverified. The paper also does not report variance, error bars, or statistical tests, so the small reported differences between configurations (39.6, 39.8, 42) may not be meaningful. The dataset itself is not yet publicly released, only promised as an 'improved version,' which limits immediate reproducibility.","major_comments":[{"comment":"The ground-truth generation pipeline is the load-bearing assumption of the entire evaluation, yet no validation is reported. The masks are produced by parsing existing 2D bounding boxes, selecting LiDAR points inside those boxes, classifying the points as moving or static, projecting them onto the fisheye image, and taking a convex hull. There is no manual verification, no inter-annotator agreement, and no sensitivity analysis of the LiDAR moving/static classifier or the ego-motion compensation. Because moving pixels constitute only 0.54% of a frame on average, even a small false-positive rate among static pixels can dominate the reported MOD IoU. The paper should validate the annotation pipeline against a manually labeled subset and report precision/recall of the generated masks, or at minimum provide a qualitative error analysis.","section":"Section 3, Semi-automated annotation procedure"},{"comment":"The central claim that rectilinear models do not generalize to fisheye images is confounded by dataset size and scene domain. Row 1 trains on 1300 KITTI frames, row 2 trains on 3638 fisheye parking frames; the 10% versus 39.8% IoU gap could be caused by the amount of training data, the scene distribution, or object classes rather than by the camera model. A cleaner control would train the same architecture on an equal number of fisheye frames, or fine-tune the KITTI model on a matched fisheye subset. Without such a control, the conclusion 'models learnt on rectilinear cameras do not generalize well to fisheye cameras' is not established by these numbers.","section":"Table 1, rows 1-2"},{"comment":"No error bars, standard deviations, or statistical tests are reported for any of the IoU values. Given that the moving-object class is only 0.54% of the pixels, the mIoU of 69.5 is dominated by the static class, and the informative quantity is the MOD IoU, whose values (39.6, 39.8, 42) differ by only a few points. The paper should report results over multiple training runs or test splits to show that the observed improvements from augmentation are not within run-to-run noise.","section":"Section 4, Results and Table 1"},{"comment":"The abstract states 'accuracy of 40% IoU and 69.5% mIoU,' but Table 1 reports MOD IoU of 39.8 for the corresponding row and 42 for the augmented row. The headline number should match the table exactly, and the paper should clarify which configuration the abstract refers to. This inconsistency matters because the 40% IoU is the paper's most prominent quantitative claim.","section":"Abstract and Table 1"}],"minor_comments":[{"comment":"The text says the fisheye network was trained on '3k fisheye images,' while Section 3 reports 3638 training frames; please make the numbers consistent.","section":"Section 4, Experimental Setup"},{"comment":"The weighted cross-entropy loss is mentioned twice, but the class weights are never specified. Reporting the exact weights would improve reproducibility.","section":"Section 4, Proposed Model"},{"comment":"The 'static objects scene augmentation' row increases the training set from 3638 to 5849 samples, but the augmentation procedure is described only qualitatively. Please specify how static scenes were selected and whether the moving-object annotations were unchanged.","section":"Section 4, Results"},{"comment":"The embedded platform is described only as '1 teraflops automotive embedded system.' Please name the hardware and state whether the 15 fps measurement includes both encoder passes or only the reused-encoder inference path.","section":"Section 4, Results"},{"comment":"The paper promises to release an 'improved version' of the dataset but does not state when or under what license, and the current version is not available. A statement on data availability would strengthen the reproducibility claims.","section":"Section 3, Dataset Statistics"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FisheyeMODNet is worth knowing about mainly because it introduces the first public fisheye MOD dataset (5139 frames, parking scenes, 0.54% moving pixels) and reports a concrete transfer gap: a rectilinear MODNet drops to 10% IoU on fisheye, while their fisheye-trained model gets about 40% IoU and 69.5% mIoU at 15 fps on a 1 TOPS embedded system. That transfer gap is the meaningful new empirical result, and it is plausible. The architecture itself is ShuffleSeg with a two-stream shared encoder from MODNet; that is incremental, not a new mechanism, and the paper does not oversell it.\n\nWhat the paper does well: it identifies a real gap in the literature — public datasets for surround-view fisheye MOD — and it honestly compares training-domain choices and augmentation/static-scene variants. The 42% IoU after balanced augmentation is a sensible baseline number for later work. They also document the main failure mode (static pedestrians detected as moving), which is the kind of detail that helps future authors.\n\nThe soft spots are exactly where the stress test lands. First and most important, the ground truth comes from a semi-automatic LiDAR pipeline (parse 2D boxes, select points, classify moving/static, project, convex hull) with no reported manual verification or sensitivity analysis. Because moving pixels are only 0.54% of a frame, a small false-positive rate among static pixels will dominate MOD IoU. Every number in Table 1 inherits whatever error that pipeline has. That is a load-bearing unvalidated assumption, not a minor nit. Second, the KITTI-vs-fisheye comparison (row 1 vs 2) is confounded: 1,300 KITTI samples vs 3,638 fisheye samples, and different scene types (parking vs driving). So the 10 vs 40 difference does not isolate camera geometry. A same-size, same-domain comparison would be needed to support the claim that rectilinear models fail specifically because of fisheye distortion. Third, smaller but real: the abstract says 40% IoU while Table 1 reports 39.8, and there are no error bars or statistical tests; given the tiny moving-pixel fraction, variance across scenes is likely meaningful. Finally, data and code are not released at publication time; the paper promises an \"improved version\" under WoodScape, so reproducibility is currently limited.\n\nNone of this kills the paper. The dataset contribution and the transfer-gap finding are worth taking seriously, but the numerical claims should be treated as provisional until the annotation pipeline is validated or the masks are released for inspection. The citation pattern is standard for this group; pointing to WoodScape as the dataset release venue is expected.\n\nWho this is for: people working on surround-view fisheye perception, parking/urban MOD, or efficient embedded segmentation. It deserves a serious referee, but the referee should demand an annotation-quality study and a deconfounded transfer experiment before final acceptance. My recommendation: send it to review with heavy revision required, not a desk reject.","headline":"A useful first benchmark for fisheye moving-object detection, but every accuracy number sits on an unvalidated semi-automatic annotation pipeline and the KITTI comparison is confounded.","tokens_in":5932,"tokens_out":2935,"would_cite":true,"duration_ms":24467,"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 lightweight two-stream CNN trained on fisheye surround-view images detects moving objects at embedded speed, lifting moving-object IoU from the 10% of a rectilinear-trained model to about 40%.","keywords":["moving object detection","fisheye cameras","surround-view","autonomous driving","semantic segmentation","two-stream CNN","embedded systems","LiDAR annotation"],"falsifier":"Have human annotators redraw ground-truth moving-object masks on a randomly chosen subset of the test frames and compare them to the semi-automated convex-hull labels; if the agreement is low (say, label IoU below roughly 0.8) or if most disagreements involve static pedestrians labelled as moving, then the exact 40% versus 10% IoU comparison and the augmentation gains are not trustworthy as reported.","tokens_in":4906,"feed_emoji":"🚗","tokens_out":12148,"duration_ms":97208,"temperature":0.7,"pith_summary":"The paper sets out to show that moving object detection—deciding which visible objects around a vehicle are actually in motion—can be done directly on 360-degree fisheye surround-view images by a small CNN fast enough for embedded automotive hardware. It introduces a new dataset of fisheye parking-scene frames with moving-object masks and a two-stream network that takes two consecutive frames and learns motion without optical-flow preprocessing. The same architecture trained on ordinary rectilinear images reaches only 10% IoU (Intersection over Union, the standard overlap score) on fisheye test frames, while training on fisheye data raises moving-object IoU to about 40% and mean IoU to 69.5%, at 15 fps on a 1-teraflops embedded system. If the results hold, surround-view fisheye systems can support near-field collision-risk estimation in real time, and rectilinear-trained moving-object detection models cannot simply be ported to fisheye cameras.","feed_headline":"Fisheye training lifts moving-object IoU from 10% to 40%","feed_subtitle":"Lightweight two-frame CNN brings 360-degree moving-object detection to embedded hardware.","key_machinery":"The load-bearing mechanism is a two-stream encoder-decoder with shared weights: two consecutive fisheye frames pass through the same lightweight encoder built from point-wise group convolutions and channel shuffling, and a three-deconvolution-layer decoder outputs a binary moving/static mask. Because the input is the frame pair itself, motion is encoded directly in the two streams and no optical-flow preprocessing or camera-specific flow distortion correction is needed. The ground-truth generation pipeline is the other supporting mechanism: LiDAR points falling inside existing object bounding boxes are classified as moving or static, projected to the image via calibration, and wrapped into convex-hull polygons; weighted cross-entropy handles the extreme class imbalance, with about 0.54% moving pixels per frame.","core_discovery":"The central discovery is that fisheye-specific training, not generic motion cues, is what unlocks moving-object detection on surround-view cameras. Feeding two temporally sequential fisheye images into a weight-shared two-stream encoder lets the network separate ego-motion from true object motion implicitly, and this model trained on fisheye data reaches 39.8% moving-object IoU and 69.5% mIoU, whereas the same two-stream design trained on rectilinear automotive data collapses to 10% IoU when evaluated on fisheye frames. Weight sharing between the two encoders costs almost nothing (69.5% mIoU, 39.6% IoU), and augmenting with annotation-free static-only scenes nudges IoU to 42%. The authors frame this as evidence that non-linear fisheye projection geometry changes the motion pattern enough that domain-specific data is required.","pith_inferences":["Beyond the paper: the rectilinear-versus-fisheye gap likely mixes projection geometry with dataset domain shift (road scenes vs. parking scenes), since the two training sets differ in both; a controlled experiment would fine-tune the rectilinear model on a small fisheye set to isolate the cause.","Beyond the paper: the convex-hull labels will under- or over-segment objects whose LiDAR points are sparse or occluded, so the reported numbers should be checked against manually drawn masks on a sample before the exact gains are taken at face value.","Beyond the paper: because the two encoders share weights, a deployable system could cache the previous frame's encoded features and run inference at higher than 15 fps with only one encoder pass per new frame, an extension the paper leaves unexplored.","Beyond the paper: although the network is trained on parking scenes, the same two-frame design may transfer to other fisheye settings such as surveillance; the paper's own observation that static pedestrians are the main false-positive source hints that appearance cues leak into the motion decision, so scene-agnostic generalization should be tested explicitly."],"forward_implications":["Surround-view fisheye camera systems can run moving-object detection on embedded hardware at 15 fps without a separate optical-flow computation stage.","Moving-object detection models trained on rectilinear camera data should not be expected to transfer to fisheye cameras; fisheye-specific training data is needed.","Sharing encoder weights across the two time steps costs less than a point of IoU (39.8 to 39.6), so deployment can reuse the previous frame's features from one inference to the next.","Adding annotation-free static-only scenes raises IoU from roughly 40% to 42%, offering a cheap way to grow the training set.","The public release of the fisheye moving-object dataset gives the community a common benchmark for comparing moving-object detection methods on near-field surround-view cameras."],"supporting_citations":[{"why":"Supplies the rectilinear-trained baseline and dataset; training the two-stream design on it gives the 10% IoU result on fisheye frames that motivates fisheye-specific training.","marker":"[14]"},{"why":"Provides the lightweight segmentation base network that the two-stream architecture adapts for the moving-object task.","marker":"[5]"},{"why":"Supplies the encoder design using point-wise group convolutions and channel shuffling, plus the idea of combining motion and appearance streams.","marker":"[11]"},{"why":"Defines the sampling strategy used to select the 5,139 fisheye frames from the capture sequences.","marker":"[16]"},{"why":"Describes the surround-view fisheye dataset program under which the new moving-object annotations are to be released.","marker":"[18]"},{"why":"Figures as a prior two-stream motion-plus-appearance segmentation method that the sequential-image training approach builds on.","marker":"[8]"}],"fun_headline_variants":["Fisheye-specific training lifts moving-object IoU to 40%","Weight-shared fisheye CNN hits 40% IoU on moving objects","Surround-view MOD: fisheye data beats rectilinear for moving objects","Fisheye MOD at 15 fps on embedded: 40% IoU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the semi-automated LiDAR-derived convex-hull masks are reliable ground truth; no manual verification or noise analysis is reported, and if those masks are systematically wrong, the reported IoU numbers and the claimed advantage of fisheye training collapse.","fun_headline_variants_meta":{"raw":{"variants":["Fisheye-specific training lifts moving-object IoU to 40%","Weight-shared fisheye CNN hits 40% IoU on moving objects","Surround-view MOD: fisheye data beats rectilinear for moving objects","Fisheye MOD at 15 fps on embedded: 40% IoU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3760,"prompt_tokens":911,"completion_tokens":2849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2763}},"tokens_in":527,"tokens_out":2849,"duration_ms":18566,"temperature":1.0,"reasoning_tokens":2763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:06:18.149913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have human annotators redraw ground-truth moving-object masks on a randomly chosen subset of the test frames and compare them to the semi-automated convex-hull labels; if the agreement is low (say, label IoU below roughly 0.8) or if most disagreements involve static pedestrians labelled as moving, then the exact 40% versus 10% IoU comparison and the augmentation gains are not trustworthy as reported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rectilinear-trained baseline and dataset; training the two-stream design on it gives the 10% IoU result on fisheye frames that motivates fisheye-specific training."},{"cited_title":"ShuffleSeg: Real-time Semantic Segmentation Network","cited_arxiv_id":"1803.03816","evidence_quote":"Provides the lightweight segmentation base network that the two-stream architecture adapts for the moving-object task."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the encoder design using point-wise group convolutions and channel shuffling, plus the idea of combining motion and appearance streams."},{"cited_title":"Uric ´ar, D","cited_arxiv_id":null,"evidence_quote":"Defines the sampling strategy used to select the 5,139 fisheye frames from the capture sequences."}],"review_version":1}