{"id":"f5904118-cf24-4f26-8ca3-661b60939604","arxiv_id":"2501.05997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Applying realistic camera occlusions to nuScenes images drops BEV vehicle segmentation IoU from 47.4 to 34.3 for cameras alone, while fusing LiDAR and radar cuts the degradation from 27.6% to 15.5%.","lead":"This paper measures how dirt and rain on car cameras degrade bird's-eye-view vehicle segmentation, then shows that adding radar and LiDAR data recovers most of the lost accuracy. It is a benchmark-style result for engineers building sensor-fusion perception systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline mitigation numbers in Table 1 rest on an unvalidated synthetic occlusion protocol — WoodScape binarized masks blurred with a 251×251 Gaussian on nuScenes — so the reported 27.6%→15.5% degradation reduction is not shown to transfer to real dirt, rain, or fog.","rationale":"The paper's contribution is an empirical measurement, not a new method, so its value depends on whether the measured degradation and mitigation transfer to the real-world conditions named in the title and abstract. The only support for that transfer is the synthetic occlusion procedure in the Dataset section. That procedure is unvalidated: WoodScape is a fisheye dataset, nuScenes cameras are narrow-FOV, and the 251×251 Gaussian kernel is a large, arbitrary choice that turns 'occlusion' into a blurred copy of the scene rather than an opaque or textured lens contaminant. Clean radar and LiDAR inputs then trivially compensate because they are not degraded, making the fusion result expected rather than a robust demonstration. This concern is not an internal contradiction, but it is a correctness risk for the central quantitative claim. The direction of the effect is plausible, and random-box occlusion also points the same way, so rejection is not warranted; the appropriate verdict remains conditional pending validation of the occlusion realism. The reader's weakest assumption identifies the same issue, and no further verdict adjustment is needed.","tokens_in":8267,"tokens_out":9158,"duration_ms":99844,"concrete_test":"Evaluate the same trained Simple-BEV C-only and C+R+L models on the same nuScenes samples under two occlusion conditions: (1) the paper's WoodScape-binary + 251×251 Gaussian-blur protocol, and (2) a physically realistic lens-soiling render, e.g., alpha-compositing actual WoodScape soiled patches, rendering fog or raindrops, or using real rainy nuScenes frames, with comparable spatial extent. If the camera-only IoU drop (47.4→34.3) and the C+R+L IoU (64.5→54.5) are not reproduced within a few IoU points under condition (2), the quantitative central claim is an artifact of the blur protocol rather than a measure of realistic occlusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assertion is the 'Realistic occlusion' row of Table 1 (C: 34.3 IoU vs C+R+L: 54.5; degradation 27.6% vs 15.5%). For this to support the central claim, the synthetic occlusion must behave like real lens soiling. The Dataset section / Fig. 2 describes taking WoodScape soiling patterns, converting to binary masks, and applying a 251×251 Gaussian blur 'only the occluded sections.' This is neither calibrated nor validated. WoodScape images are fisheye; nuScenes cameras are not, and no geometric or photometric remapping is described. More importantly, a blurred copy of the original image is not opaque dirt or raindrop occlusion: scene content remains visible, and the amount of information lost depends on an arbitrary kernel size. Since radar and LiDAR are always clean, any strong image corruption will make fusion look beneficial; the specific '15.5% residual' is therefore a property of the blur protocol. The paper also gives no comparison to real occluded or rainy frames, no code, and no error bars, so this is not merely a missing statistical nicety but a threat to the external validity of the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how camera lens occlusion affects bird's-eye-view (BEV) vehicle segmentation and whether multi-sensor fusion with radar and LiDAR mitigates the degradation. The authors start from the Simple-BEV architecture on nuScenes and create three occlusion conditions: random boxes, occlusion over overlapping camera fields of view, and a so-called \"realistic\" occlusion obtained by binarizing soiling patterns from the WoodScape dataset and blurring the occluded regions with a 251x251 Gaussian kernel. They report IoU values for camera-only (C), camera+radar (C+R), camera+LiDAR (C+L), and camera+radar+LiDAR (C+R+L) under each condition, with the central quantitative result in Table 1: under \"realistic\" occlusion, IoU drops to 34.3 for C but only to 54.5 for C+R+L, and the relative degradation decreases from 27.6% to 15.5%. The paper concludes that fusing clean radar and LiDAR data with occluded cameras substantially restores BEV segmentation accuracy.","tokens_in":8578,"tokens_out":4197,"duration_ms":43043,"significance":"If the result holds, the paper provides a useful empirical measurement of how sensor fusion can compensate for camera soiling in BEV perception, which is relevant for autonomous driving robustness. The authors use public datasets (nuScenes, WoodScape) and a public architecture (Simple-BEV), and the core numbers in Table 1 are direct measurements rather than fitted quantities, which is a strength. However, the current evidence is preliminary: the synthetic occlusion protocol is not validated against real dirt, rain, or fog; the baseline IoU values for C, C+R, and C+L are quoted from the Simple-BEV paper rather than retrained under the same protocol; and no error bars, repeated runs, or statistical tests are provided. These issues directly affect the external validity and reliability of the headline claim, so the contribution is not yet established at the level expected for a journal publication. The paper is a reasonable starting point for a more rigorous empirical study, but it currently reads more like a short report or workshop contribution than a complete journal article.","major_comments":[{"comment":"The \"realistic\" occlusion protocol is not validated, and the specific numbers in Table 1 depend on an uncalibrated 251x251 Gaussian blur applied to binarized WoodScape patterns. WoodScape images are fisheye while nuScenes cameras are not, and the paper does not describe any geometric or photometric remapping. More importantly, a blurred copy of the original image content is not equivalent to opaque dirt, raindrops, or fog: scene information remains partially visible, and the amount of information loss is an artifact of the arbitrary kernel size. Since radar and LiDAR inputs are always clean, sufficiently strong image corruption will make fusion appear beneficial, so the 27.6% to 15.5% degradation reduction is not shown to transfer to real-world soiling. The paper should validate the protocol against real occluded or rainy frames (e.g., nuScenes rain sequences or WoodScape soiled images) or at least include a sensitivity analysis over kernel size and blur opacity, and ideally provide real-occlusion test results.","section":"Quantitative Analysis, Table 1"},{"comment":"The baseline values for C, C+R, and C+R+L are not obtained under a common protocol. The text states that C, C+R, and C+L IoU values (47.4, 55.7, 60.8) are reported from the Simple-BEV paper, while C+R+L (64.5) was trained by the authors because pretrained weights were not available. The degradation percentages in Table 1 are computed against this mixed set of baselines, so the comparison is only meaningful if the quoted Simple-BEV numbers are directly comparable to the authors' training setup (80 epochs, batch size 12, input resolution 448x800) and evaluation protocol. The authors should retrain all four configurations under identical conditions, or clearly justify the transferability of the quoted numbers. Otherwise, even a monotonic trend in Table 1 could be an artifact of differing training protocols rather than of fusion.","section":"Quantitative Analysis, Table 1, Implementation Details"},{"comment":"All IoU values in Table 1 are single numbers with no error bars, no repeated runs, and no significance tests. The conclusion relies on relatively small differences among conditions, e.g., the realistic condition C+R+L at 54.5 versus C+L at 50.3, and degradation percentages of 15.5% versus 17.2%. Without variance estimates or a stated number of evaluation samples, these differences may be within run-to-run noise, especially given that the self-trained C+R+L model could have a different random seed or data split from the quoted baselines. The authors should report mean and standard deviation over multiple runs (or at least over evaluation subsets) and state the number of test scenes used.","section":"Table 1, Quantitative Analysis"},{"comment":"The experimental protocol is under-specified, which makes the results difficult to reproduce or interpret. The paper does not state whether the nuScenes training or validation split is used for occlusion evaluation, how many samples are occluded and evaluated, whether all six cameras are occluded simultaneously, how patterns are selected and placed per image, and crucially whether the models are trained on clean images and tested on occluded images, or trained with occlusion augmentation. The text says \"we create an occluded version of the nuScenes dataset\" but also describes training a C+R+L baseline, so the reader cannot tell if the occlusion is applied only at test time. The authors should provide a complete protocol description, including the exact evaluation split and sample count.","section":"Dataset section, Implementation Details"}],"minor_comments":[{"comment":"The abstract contains the raw YouTube URL \"https://youtu.be/OmX2NEeOzAE\" in the text body; this should be moved to a footnote or removed.","section":"Abstract"},{"comment":"The paper uses inconsistent capitalization and spelling for the same terms, e.g., \"Woodscape\" versus \"WoodScape\" and \"simple-BEV\" versus \"Simple-BEV\". Please standardize these names.","section":"Throughout"},{"comment":"The BCEWithLogitsLoss formula in Equation (1) is written with a sum over an unspecified N; please clarify whether the average is over pixels or over samples, and ensure the notation matches the implementation referenced in [15].","section":"Equation (1)"},{"comment":"The caption says the third column shows \"a binary mask indicating the areas of occlusion in red,\" but the main text describes the third column as a BEV projection; please align the column descriptions with what is actually displayed.","section":"Figure 4 caption"},{"comment":"The first column header \"Baseline Architecture [15]\" is misleading because the rows are occlusion conditions, not baselines. Consider renaming it to \"Occlusion Condition\" and using the first row as \"No occlusion (Simple-BEV).\"","section":"Table 1"},{"comment":"The reference for the soiling patterns is given as [17], the general WoodScape dataset paper, but the text refers to the \"WoodScape Soiled Dataset.\" Please clarify whether a separate soiling dataset was used and cite it specifically if so.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest in its Future Work section about not yet addressing rain, different occlusion levels, and other sensor degradations, which supports the view that the current empirical scope is narrow. The main scientific risk is that the synthetic occlusion protocol may not represent real soiling, and the mixed-origin baselines weaken the quantitative claim. These issues are fixable within the manuscript's scope, but they require substantial additional experiments or careful reframing of the claims. I would not recommend rejection, but the paper needs more than light revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is the measured IoU table: camera occlusions degrade BEV vehicle segmentation, realistic WoodScape-based occlusion hurts most, and adding radar and LiDAR at inference time recovers a large share of the loss. That result is plausible and the paper is the first to report it for Simple-BEV on nuScenes. The C+R+L baseline trained by the authors is also a legit addition.\n\nThe soft spot is exactly what the stress-test flags: the occlusion synthesis. Taking WoodScape soiling patterns, binarizing them, and blurring the occluded regions with a 251×251 Gaussian is not validated against real dirt, rain, or fog. The kernel size is arbitrary, and a blurred copy of the image still shows scene structure, so the amount of information lost is a property of the blur radius. The headline 27.6%→15.5% degradation reduction therefore doesn't transfer to real conditions until someone shows it does. No code, no data release, no comparison to real rainy/soiled frames, and no error bars (single run per cell) make this hard to check.\n\nAlso, the baseline fairness issue is real: C, C+R, C+L are quoted from the Simple-BEV paper while C+R+L is self-trained. Differences in training protocol could shift relative degradations. The paper would be stronger if all four modalities were retrained under one protocol.\n\nOn the other hand, the paper is honest about its scope and doesn't oversell. The qualitative figures make sense, the monotonic trend across occlusion types is coherent, and the conclusion that fusion helps is robust even if the exact magnitudes aren't. I'd treat the numbers as illustrative, not as a calibrated benchmark.\n\nVerdict: conditional, leaning revise. This deserves a serious referee because the empirical question is timely and the protocol is fixable. Require sensitivity analysis over blur parameters, validation against real soiling or at least a clear statement of the synthetic limitation, error bars, and consistent retraining of baselines. Remove the stray YouTube link from the abstract.","headline":"Useful first measurement of occlusion effects on BEV segmentation, but the synthetic blur protocol is unvalidated, so the headline numbers are illustrative, not transferable.","tokens_in":9045,"tokens_out":2566,"would_cite":false,"duration_ms":24816,"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":"Fusing radar and LiDAR with occluded cameras cuts the relative degradation of bird's-eye-view vehicle segmentation under realistic lens soiling from 27.6% to 15.5%.","keywords":["multi-sensor fusion","bird's eye view","vehicle segmentation","camera occlusion","synthetic soiling","nuScenes","Simple-BEV","autonomous driving"],"falsifier":"Train or evaluate the same Simple-BEV configurations on a real soiling dataset—cameras with physically dirty or wet lenses in a controlled driving scene—and compare the camera-only and camera-radar-LiDAR IoU drops. If the relative degradation under realistic soiling is not substantially larger than the degradation from random-box occlusion, or if fusing radar and LiDAR does not reduce the relative degradation below the camera-only level, the paper's central claim would be contradicted.","tokens_in":8089,"feed_emoji":"🚗","tokens_out":9171,"duration_ms":65028,"temperature":0.7,"pith_summary":"The paper tests how camera lens occlusions—simulated dirt, moisture, and fog from the WoodScape soiling dataset applied to nuScenes multi-view images—hurt bird's-eye-view (BEV) vehicle segmentation. It reports that realistic occlusion is the most damaging of three occlusion types, dropping camera-only intersection-over-union (IoU) from 47.4 to 34.3. Fusing the occluded cameras with radar and LiDAR in the Simple-BEV architecture recovers most of the loss, raising IoU to 54.5 and cutting relative degradation from 27.6% to 15.5%. The central claim is that multi-sensor fusion is an effective practical countermeasure to lens soiling for BEV perception.","feed_headline":"Fusing radar and LiDAR cuts camera-occlusion BEV degradation to 15.5%","feed_subtitle":"In simulated dirt and fog, camera-only IoU falls to 34.3; adding radar and LiDAR holds it at 54.5.","key_machinery":"The load-bearing mechanism is the Simple-BEV fusion pipeline [15]: a ResNet-101 camera encoder extracts 2D features, bilinear sampling lifts them to a 3D BEV plane, and voxelized radar and LiDAR occupancy grids are flattened and merged into the same BEV representation before a ResNet-18 encoder and segmentation head. The other key piece is the occlusion synthesis: WoodScape [17] soiling patterns are binarized and blurred with a 251x251 Gaussian kernel to produce realistic lens occlusions on the nuScenes [16] camera views.","core_discovery":"The paper's central discovery is that realistic camera occlusion, modeled as binary WoodScape soiling masks blurred with a 251x251 Gaussian kernel, degrades BEV vehicle segmentation more than random or overlap occlusion, and that this degradation is systematically mitigated by adding radar and LiDAR. In the camera-only baseline, IoU falls to 34.3; with camera, radar, and LiDAR fused it stays at 54.5, so the relative degradation falls from 27.6% to 15.5%. The result is presented as evidence that complementary sensors compensate for occluded camera views in the BEV domain.","pith_inferences":["The same fusion approach would likely reduce occlusion-driven degradation for other BEV tasks such as object detection and lane segmentation, since the BEV representation is shared; this extrapolates beyond the paper's vehicle-segmentation experiments.","The degradation numbers may underestimate real-world soiling because the synthesis uses static binary masks and does not model glare, transparent droplets, or temporal accretion; real-world field tests could show different magnitudes.","The monotonic improvement with added sensors suggests that even radar-only fusion is a worthwhile fallback when LiDAR is unavailable, a cost-benefit conclusion the paper does not draw."],"forward_implications":["Under realistic camera soiling, BEV vehicle segmentation degrades substantially; camera-only IoU drops by 27.6%, so systems relying on cameras alone need a mitigation strategy.","Adding radar alone reduces the relative degradation to 22.6%, adding LiDAR alone to 17.2%, and adding both to 15.5%, so each extra modality contributes measurably.","The fusion benefit holds for all three occlusion types tested, with the largest absolute gains on the hardest, realistic occlusion.","The pattern suggests that multi-sensor fusion is a general robustness mechanism for BEV perception under lens contamination, not just for the specific soiling patterns used in the paper."],"supporting_citations":[{"why":"Supplies the Simple-BEV fusion architecture and the baseline IoU values (camera-only 47.4, camera+radar 55.7, camera+LiDAR 60.8) that all occlusion experiments are compared against.","marker":"[15]"},{"why":"Provides the nuScenes multi-view camera, radar, and LiDAR data and the vehicle segmentation labels used for training and evaluation.","marker":"[16]"},{"why":"Supplies the WoodScape soiling patterns that the paper converts into binary masks and Gaussian-blurred occlusions to simulate realistic lens dirt, moisture, and fog.","marker":"[17]"}],"fun_headline_variants":["Fusing radar and LiDAR cuts camera-occlusion BEV loss to 15.5%","Multi-sensor fusion reduces BEV occlusion degradation by 44%","Radar and LiDAR compensate for occluded cameras in BEV perception","Camera-only BEV drops to 34.3 IoU; adding radar and LiDAR holds 54.5"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole measurement of degradation and mitigation assumes that WoodScape soiling patterns, converted to binary masks and blurred with a 251x251 Gaussian kernel, faithfully reproduce real-world lens occlusion from dirt, rain, and fog.","fun_headline_variants_meta":{"raw":{"variants":["Fusing radar and LiDAR cuts camera-occlusion BEV loss to 15.5%","Multi-sensor fusion reduces BEV occlusion degradation by 44%","Radar and LiDAR compensate for occluded cameras in BEV perception","Camera-only BEV drops to 34.3 IoU; adding radar and LiDAR holds 54.5"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000451,"raw_usage":{"total_tokens":2242,"prompt_tokens":886,"completion_tokens":1356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1263}},"tokens_in":502,"tokens_out":1356,"duration_ms":11311,"temperature":1.0,"reasoning_tokens":1263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:49.722969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or evaluate the same Simple-BEV configurations on a real soiling dataset—cameras with physically dirty or wet lenses in a controlled driving scene—and compare the camera-only and camera-radar-LiDAR IoU drops. If the relative degradation under realistic soiling is not substantially larger than the degradation from random-box occlusion, or if fusing radar and LiDAR does not reduce the relative degradation below the camera-only level, the paper's central claim would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Simple-BEV fusion architecture and the baseline IoU values (camera-only 47.4, camera+radar 55.7, camera+LiDAR 60.8) that all occlusion experiments are compared against."},{"cited_title":"Caesar, V","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes multi-view camera, radar, and LiDAR data and the vehicle segmentation labels used for training and evaluation."},{"cited_title":"Yogamani, C","cited_arxiv_id":null,"evidence_quote":"Supplies the WoodScape soiling patterns that the paper converts into binary masks and Gaussian-blurred occlusions to simulate realistic lens dirt, moisture, and fog."}],"review_version":1}