{"id":"4dd5d884-fc4f-43b3-b533-0d4db1fe40ee","arxiv_id":"1909.00532","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Panoramic training images cut from 360-degree synthetic scenes, especially 180-degree FoV views, improve semantic segmentation accuracy and distortion robustness compared with conventional training data.","lead":"The authors stitch four-direction synthetic street images into 360-degree panoramas, creating a new dataset called SYNTHIA-PANO, and train a segmentation network on it. They report that training on 180-degree panoramic views improves accuracy and distortion resistance compared with training on ordinary forward-view images.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's anti-distortion claim is unsupported because Sec. 4.4 and Fig. 12 contradict each other about which model wins for focal lengths below 625.","rationale":"The paper's central assertion is that training on stitched SYNTHIA-PANO panoramas improves both segmentation quality and distortion robustness relative to training on original four-direction images. The synthetic-to-real transferability concern raised by the reader is real, but it is not the most load-bearing issue here: the paper's own in-domain evaluation for the distortion-robustness component is internally inconsistent. A claim cannot be validated on real data if the reported synthetic experiment is self-contradictory. I therefore focus on Sec. 4.4. The three mutually inconsistent statements (two in the text, one in the Fig. 12 caption) make it impossible to determine from the manuscript whether Mpano180 outperforms Morin under strong distortion. Since this experiment is the sole support for the abstract's 'better capacity to resist the image distortion' claim, that part of the central claim is unsupported as written. I do not treat the contradiction as evidence of misconduct; a typographical or reporting error is plausible. But it is exactly the kind of error that must be corrected and re-verified before the conclusion is accepted. The reader's own rationale noted this contradiction, even though their formal weakest assumption was the synthetic-to-real proxy; I list partial agreement for that reason. The proposed test is a clean rerun with per-focal-length numbers and error bars, or at minimum a corrected figure/text consistent with the raw data. With the current text, the REJECT verdict stands: the dataset may still be useful, but the paper's headline claims are not supported by the reported experiments.","tokens_in":11971,"tokens_out":8018,"duration_ms":274848,"concrete_test":"Recompute the Fig. 12 experiment cleanly: evaluate Morin and Mpano180 on the same Seqs05-summer test panoramas at focal lengths f=700, 625, 600, 500, and 400, using at least three training seeds, and report per-f mIoU with the underlying raw values. Then check the sign of the difference Mpano180 - Morin at each f; if the difference is not positive for f<625 (or is negative throughout), the distortion-robustness claim should be removed from the abstract and conclusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.4 (Evaluation of Anti-distortion Ability) contains mutually incompatible statements about the same experiment. It says, \"When the focal length is shorter than 625, Morin performs better than Mpano180,\" and then, two sentences later, \"When the focal length is shorter than 625, the model Mpano180 trained with panoramic images always preforms better, which means that the anti-distortion ability of the model is strengthened.\" The caption of Fig. 12 adds a third version: \"When the focal length is in 700 to 400, the mIoU of Morin is higher than that of Mpano180.\" If Fig. 12's caption is correct, Mpano180 never outperforms Morin in the tested range and the abstract's claim of improved distortion robustness is false. If the intended result is that Mpano180 wins only for f<625, then the text and caption are still inconsistent and no reliable conclusion can be drawn without corrected numbers. This contradiction is internal to the paper, not a disagreement with external consensus, and it directly undermines one of the three central claims in the abstract.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SYNTHIA-PANO, a synthetic panoramic semantic segmentation dataset built by cylindrical projection and stitching of the four fixed 100-degree views in SYNTHIA-Seqs, and studies how training data view direction and FoV affect ICNet segmentation performance. The authors compare models trained on original four-direction images versus panoramic crops with 90, 180, and 360 degree FoVs, and report that 180-degree FoV training yields the best accuracy and mIoU on a panoramic test sequence. The paper further claims that the model trained on panoramic images has better anti-distortion ability, evaluated by feeding both models cylindrically projected images with focal lengths from 400 to 700.","tokens_in":12141,"tokens_out":6150,"duration_ms":60715,"significance":"If the empirical claims were reliable, the released dataset and simple region-matching stitching method would be a useful resource for studying panoramic semantic segmentation with standard architectures. The FoV comparison is partly controlled by fixing epoch and iteration counts and by feeding the same image content across batch sizes, and the authors ship code and dataset artifacts. However, the central claims are entirely empirical and are supported by a single synthetic test sequence with no variance reporting. The anti-distortion claim is internally contradicted by the paper's own text and figure caption, and the 180-degree optimum is selected on the same evaluation set used for final numbers. The practical relevance for real panoramic imagery is acknowledged by the authors as an open gap, which further limits the scope of the conclusions.","major_comments":[{"comment":"The anti-distortion experiment is presented with mutually incompatible conclusions. The text says, \"When the focal length is shorter than 625, Morin performs better than Mpano180,\" and then says, \"When the focal length is shorter than 625, the model Mpano180 trained with panoramic images always preforms better,\" while the caption of Fig. 12 says, \"When the focal length is in 700 to 400, the mIoU of Morin is higher than that of Mpano180.\" These statements cannot all be true, and the abstract's third central claim about improved distortion robustness depends on Mpano180 outperforming Morin under severe distortion. This internal contradiction leaves the anti-distortion conclusion unsupported as written; the authors must reconcile the text, caption, and underlying numbers before the claim can be evaluated.","section":"Sec. 4.4 and Fig. 12"},{"comment":"The claim that 180-degree FoV is the best training choice is based on evaluation on a single sequence (Seqs05-summer) with no repeated seeds, no error bars, and no statistical comparison. The same evaluation set is used both to select the best FoV and to report the final numbers, so the result is vulnerable to selection on the test set. A controlled comparison with multiple training runs, mean and standard deviation, and a held-out validation set is needed to support the central \"180-degree is best\" finding.","section":"Sec. 4.3, Fig. 10, Table 4"},{"comment":"The inverse cylindrical projection in Eq. (2) does not appear to be the inverse of the forward mapping in Eq. (1). With r = f, Eq. (1) gives x' = f * atan(x/f), whose inverse is x = f * tan(x'/f); Eq. (2) instead states x = f * atan(x'/r), which is not the inverse. If this equation reflects the implemented mapping, the stitching pipeline may be incorrect. The authors should verify the forward/backward pair and report the correct equations.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The abstract and conclusion generalize the findings to \"panoramic images\" and to practical applications such as autonomous driving, but all training and evaluation are performed on synthetic SYNTHIA-PANO images built from the same SYNTHIA source used for training. The paper itself acknowledges in the conclusion that the gap between synthetic and real data is a key limiting factor. The representativeness of these cylindrical synthetic panoramas for real panoramic imagery, including real lens distortion and stitching artifacts, is untested, so the practical conclusions overstate the evidence.","section":"Sec. 5 and Abstract"}],"minor_comments":[{"comment":"The discrepancy measure in Eq. (7) sums over p in R1 and q in R2 without specifying a correspondence; presumably it computes the sum of absolute differences over aligned pixel positions, and this should be stated explicitly.","section":"Sec. 3.3, Eq. (7)"},{"comment":"The statement that \"the order of the directions is randomly chosen\" is unclear: if the directional order varies across panoramas, the fixed transformation structure in Eq. (3) and the equal spacing assumption in Eq. (5) need to be reconciled.","section":"Sec. 3.4"},{"comment":"The learning-rate decay policy is described only as \"learning rate decay policy is used\"; the specific schedule and decay parameters should be reported for reproducibility.","section":"Sec. 4.1"},{"comment":"There are minor typos: \"orin\" should be \"origin\" in Table 4, and \"preforms\" should be \"performs\" in Sec. 4.4.","section":"Table 4 and Sec. 4.4"},{"comment":"There are several typographical errors, including \"SYHTHIA dateset\" for \"SYNTHIA dataset\" and \"a new dataset of panoramic image\" in the abstract; these should be corrected.","section":"Sec. 1"},{"comment":"Reference [28] lists \"Shum, H.-Y.\" three times in the author list and should be cleaned up.","section":"References"}],"recommendation":"reject","confidential_remarks":"The internal contradiction in Sec. 4.4 is not a matter of wording: it directly undermines one of the three central claims in the abstract, and no reliable conclusion can be drawn from the anti-distortion experiment as reported. The FoV comparison also lacks the statistical support expected for a central empirical claim. The dataset and code release are potentially useful, but the manuscript in its current form would require new experiments and a substantial rewrite; I would not recommend revision on the present evidence, though a corrected resubmission might be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful dataset paper with a clean empirical finding about 180-degree training FoV, but the anti-distortion experiment is internally contradictory and the paper needs major revision before I'd trust its conclusions.\n\nWhat is actually new: SYNTHIA-PANO, a stitched panoramic version of SYNTHIA built with cylindrical projection and a simple region-matching method to estimate overlap. The region-matching trick is a sensible response to the thin 10-degree overlaps that break ORB matching. The 180-degree FoV result is genuinely interesting: training ICNet on 180-degree crops beats both 90/360 and the original 100-degree images, and the authors honestly report the non-obvious finding that 360-degree training hurts. The training comparison is controlled for content and iteration count, which is more than many similar papers do.\n\nThe soft spots are real. The biggest is Sec. 4.4: the text first says \"When the focal length is shorter than 625, Morin performs better than Mpano180,\" then two sentences later says \"Mpano180 trained with panoramic images always performs better\" in the same regime, and the Fig. 12 caption says Morin is higher throughout 700-400. These are mutually incompatible statements about the same experiment, and the abstract's anti-distortion claim rests on that experiment. As written, that claim is unsupported. The FoV comparison also has no error bars or repeated seeds, and the 180-degree choice is selected on the same Seqs05 evaluation set used for the final numbers, so selection bias is not excluded. Equation 2 contains an inverse-mapping typo: it should be f*tan(x'/r), not f*atan(x'/r). Finally, the synthetic-to-real gap is acknowledged but never addressed; all results are on synthetic panoramas, so the practical claims are limited.\n\nNone of this is fatal to the dataset itself. The flaws are fixable, and the 180-degree finding deserves a proper test with error bars and a held-out model selection set. I'd send this to a serious referee, but I would not accept it in current form. The authors need to fix the Sec. 4.4 contradiction, correct Eq. 2, and ideally show something on real panoramic images. If that happens, this could be a modest but solid dataset contribution.","headline":"Useful dataset and an interesting 180-degree FoV result, but the anti-distortion experiment is internally contradictory and the paper needs major revision before its central claims can be trusted.","tokens_in":12701,"tokens_out":3452,"would_cite":true,"duration_ms":32212,"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":"Training a real-time segmentation network on stitched panoramic images—especially 180-degree crops—improves accuracy and distortion resistance on panoramic test imagery compared with training on ordinary directional views.","keywords":["semantic segmentation","panoramic images","synthetic dataset","SYNTHIA-PANO","cylindrical projection","image stitching","ICNet","distortion resistance"],"falsifier":"Evaluate a SYNTHIA-PANO-trained model on real panoramic street images and compare its mIoU and distortion resistance with a model trained on the original directional SYNTHIA views; if the panorama-trained model does not improve on real images, the central claim fails. A second check: use a network whose receptive field covers the whole 360-degree panorama; if 360-degree training then beats 180-degree training, the field-of-view finding is architecture-dependent.","tokens_in":11749,"feed_emoji":"🌐","tokens_out":6136,"duration_ms":54683,"temperature":0.7,"pith_summary":"The paper claims that a semantic segmentation model trained on panoramic images assembled from four directional views segments panoramic test images better than the same model trained on the original directional views. It builds a synthetic panorama dataset, SYNTHIA-PANO, by cylindrical projection and a simple region-matching stitch, then compares models trained on ordinary views, 90-degree, 180-degree, and 360-degree panoramas. The experiments show that accuracy and mean intersection-over-union rise as the training field of view grows to 180 degrees and then fall at 360 degrees, and that panorama-trained models degrade more slowly under severe cylindrical distortion. A reader would care because this offers a cheap way to obtain labeled training data for surrounding-scene perception without manual panorama annotation, provided the synthetic imagery transfers to real cameras.","feed_headline":"Stitched panoramas beat single-view training for 360° segmentation","feed_subtitle":"Training on 180-degree panoramic crops raises mIoU and improves resistance to cylindrical distortion in synthetic street scenes.","key_machinery":"The load-bearing object is the SYNTHIA-PANO dataset, produced by projecting each 100-degree direction image from SYNTHIA onto a cylinder with focal length f = 532.740352 and then stitching the four projections via horizontal translations separated by d ≈ 835 pixels, estimated by scanning overlapping regions for the minimum pixel-wise discrepancy. This yields 3340-by-760, 360-degree panoramas with pixel labels intact. The argument then runs on controlled comparisons of this data: the same ICNet architecture is trained on original views versus panoramas and on 90-degree, 180-degree, and 360-degree crops, isolating the effect of training-view geometry and field of view.","core_discovery":"The paper sets out to show that the view geometry of training data, not just the network architecture, limits panoramic semantic segmentation. Using the ICNet real-time segmenter with class-balanced training, the authors find that a model trained on four-direction SYNTHIA images mis-segments objects appearing in left or right views of a panorama, while models trained on stitched SYNTHIA-PANO images fix this defect. Among training fields of view, 180 degrees gives the best mIoU and pixel accuracy on a held-out panorama sequence, while 360 degrees underperforms it because the extreme aspect ratio and finite receptive field hurt feature extraction. The same 180-degree-panorama model also keeps higher mIoU than the baseline when test panoramas are re-projected with shorter focal lengths, meaning under stronger distortion. The claimed payoff is that panoramic training data improves both accuracy and resistance to distortion for surrounding-scene segmentation.","pith_inferences":["An untested implication: the 180-degree-over-360-degree result likely reflects ICNet's square kernels and finite receptive field, so a panorama-adapted architecture could make full 360-degree training the best choice.","A related implication: because the stitching order of directions is randomized, the model is implicitly taught to treat direction as irrelevant, which may contribute to the distortion resistance and deserves isolation as its own experimental variable.","A testable transfer: the method should be validated on real panoramic images; the paper's own conclusion acknowledges the synthetic-to-real gap, so the geometric benefit could be outweighed in practice even if it is real on synthetic data."],"forward_implications":["Panoramic training data lets an unmodified real-time segmenter handle off-forward views that appear in 360-degree sensing, so surrounding-perception systems need not rely on separately labeled directional images.","A 180-degree training field of view is a practical sweet spot for this architecture: it preserves enough spatial context to recognize whole objects while avoiding the feature dilution of the 4:1 aspect-ratio 360-degree panorama.","Panorama-trained models resist cylindrical distortion better, which matters for fisheye and annular cameras whose images are commonly unwrapped by cylindrical projection.","The stitching method transfers labels automatically from the source images, so a labeled panorama dataset can be generated from any multi-camera synthetic sequence with known overlap, avoiding manual annotation."],"supporting_citations":[{"why":"Supplies the four-direction synthetic street images and label maps that are stitched into SYNTHIA-PANO.","marker":"[5]"},{"why":"Provides the ICNet architecture used in all training and evaluation runs.","marker":"[6]"},{"why":"Provides the cylindrical projection mapping used to unify geometry before stitching.","marker":"[28]"},{"why":"Supplies the general image alignment and stitching formulation behind the translation model.","marker":"[29]"},{"why":"Is the feature matcher shown to fail on the thin overlap, motivating the proposed region-matching distance estimate.","marker":"[30]"}],"fun_headline_variants":["180° panoramas boost semantic segmentation accuracy","Panoramic training data improves distortion resistance","Synthetic panoramas sharpen semantic segmentation models","Pano training data beats single-view for segmentation","Wider FoV training improves segmentation and distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison assumes that panoramas synthetically stitched from SYNTHIA's four fixed 100-degree cameras, with a single estimated seam distance d ≈ 835, behave like real panoramic imagery well enough that conclusions about training data transfer to actual 360-degree cameras.","fun_headline_variants_meta":{"raw":{"variants":["180° panoramas boost semantic segmentation accuracy","Panoramic training data improves distortion resistance","Synthetic panoramas sharpen semantic segmentation models","Pano training data beats single-view for segmentation","Wider FoV training improves segmentation and distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000358,"raw_usage":{"total_tokens":1896,"prompt_tokens":861,"completion_tokens":1035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":967}},"tokens_in":477,"tokens_out":1035,"duration_ms":7870,"temperature":1.0,"reasoning_tokens":967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:46:32.336003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate a SYNTHIA-PANO-trained model on real panoramic street images and compare its mIoU and distortion resistance with a model trained on the original directional SYNTHIA views; if the panorama-trained model does not improve on real images, the central claim fails. A second check: use a network whose receptive field covers the whole 360-degree panorama; if 360-degree training then beats 180-degree training, the field-of-view finding is architecture-dependent.","supporting_citations":[{"cited_title":"The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,","cited_arxiv_id":null,"evidence_quote":"Supplies the four-direction synthetic street images and label maps that are stitched into SYNTHIA-PANO."},{"cited_title":"Creating full view panoramic image mosaics and environment maps,","cited_arxiv_id":null,"evidence_quote":"Provides the cylindrical projection mapping used to unify geometry before stitching."},{"cited_title":"Image alignment and stitching: A tutorial,","cited_arxiv_id":null,"evidence_quote":"Supplies the general image alignment and stitching formulation behind the translation model."},{"cited_title":"Orb: An efficient alternative to sift or surf.,","cited_arxiv_id":null,"evidence_quote":"Is the feature matcher shown to fail on the thin overlap, motivating the proposed region-matching distance estimate."}],"review_version":1}