{"id":"ea1da1d9-cbcc-45ad-8d4e-e0219bb3f328","arxiv_id":"2504.14806","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ITDNet jointly optimizes LiDAR restoration and place recognition via alternating task-driven training, and reports state-of-the-art retrieval accuracy in rain, snow, and fog.","lead":"This paper presents ITDNet, which alternates training between a LiDAR restoration network and a place-recognition network so that clean-up is guided by recognition performance. In tests on three adverse-weather datasets, ITDNet reports the highest retrieval accuracy, with the joint training beating separate clean-then-recognize pipelines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Boreas cross-season training pairs may be semantically invalid, so the only real-world evidence for ITDNet's improvement is not trustworthy.","rationale":"The reader's weakest_assumption matches the primary vulnerability I find. The paper's headline result is the performance improvement of Union over Direct and Separate; the synthetic datasets provide controlled but not real-world evidence, and Boreas is the only real-world support. The Boreas LDR pair construction (Sec. IV-A) uses pose proximity as a proxy for semantic correspondence, and the paper itself notes that environmental changes may cause misalignment. If the proxy fails for a substantial fraction of pairs, the LDR module is trained to map degraded scans toward clean scans that are not the same place, corrupting both the reconstruction loss and the LPR pseudo-label loss, and potentially teaching the restoration module to remove real scene content. The paper's assertion that the approximation is 'sufficient' because LPR improves is circular, since the same pairs generated that improvement. A concrete overlap-based validity check plus retraining on only high-confidence pairs would settle whether the real-world claim survives. I also note a secondary concern: the proposed FSS metric (Sec. IV-B) is evaluated with the same ITDNet-P module that was co-trained with ITDNet-D via the task-driven loss, so Table V's FSS comparison is partly circular; however, this does not affect the main R@1 and R@1% LPR tables, so I do not treat it as the load-bearing issue. Overall, the framework is plausible and the synthetic results are internally consistent, but the real-world evidence rests on an untested pairing assumption; CONDITIONAL remains appropriate.","tokens_in":18587,"tokens_out":8419,"duration_ms":80835,"concrete_test":"Quantify the semantic and geometric validity of the Boreas training pairs: for each generated (degraded, clean) pair, compute the fraction of projected points from the degraded scan that have a nearest neighbor in the clean scan within a small threshold (e.g., 0.2 m), restricted to static regions (ground and buildings) or using scan semantics if available. Report the distribution. Then retrain the LDR module (and full ITDNet) using only pairs above a high overlap threshold (e.g., top 70%) and re-evaluate on Boreas Seq 04 Easy/Hard. If R@1 changes by less than about 0.02, the misalignment concern is not the source of the claimed gain; if performance drops materially, the original result relied on incorrect targets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ITDNet achieves state-of-the-art place recognition in adverse weather, with Union mode outperforming Direct and Separate. Because Weather-KITTI and Weather-Apollo are synthetic, the Boreas experiment is the only real-world evidence. The LDR training pairs for Boreas (Sec. IV-A) are formed by projecting each degraded scan from Seq 02-03 onto the closest scan from Seq 00 within 0.01 m and 0.1 deg, and the paper explicitly acknowledges that 'environmental changes may lead to some misalignment.' This is the load-bearing assumption: if the paired clean scan is not the same scene (seasonal vegetation, parked cars, road surface changes), then both the reconstruction loss (Eq. 15) and the task-driven KL loss (Eq. 16) supervise the LDR module with wrong targets. The resulting LDR could learn to erase legitimate scene changes rather than weather noise, and the LPR pseudo-labels would encourage alignment to an incorrect descriptor. Since Table IV (Boreas) is the only non-synthetic demonstration, invalid pairs would undercut the real-world transfer claim even if the synthetic results are correct. The paper's defense that 'this approximation proves sufficient' is circular because it uses the improved LPR numbers as evidence for the very pairs that produced them.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ITDNet, an iterative task-driven framework that couples a LiDAR data restoration (LDR) module with a LiDAR place recognition (LPR) module. The LDR module uses Dual-Domain Mixer (DDM) and Semantic-Aware Generator (SAG) blocks, while the LPR module uses Multi-Frequency Transformer (MFT) and Wavelet Pyramid NetVLAD (WPN) blocks. Training alternates between the two modules: the LPR is trained with triplet loss, and the LDR is trained with a reconstruction loss plus a KL-divergence task-driven loss that aligns descriptors of restored scans to descriptors of clean scans using pseudo-labels from the previous LPR epoch. Experiments on Weather-KITTI, Weather-Apollo, and Boreas compare Direct, Separate, and Union modes against OverlapTransformer, LCDNet, and CVTNet. Reported results show that Union consistently outperforms Separate, which outperforms Direct, and the paper also introduces the Feature Similarity Score (FSS) metric and the Weather-Apollo dataset.","tokens_in":18858,"tokens_out":9796,"duration_ms":83843,"significance":"If the results hold, the main contribution is a practical demonstration that restoration and place recognition can be co-trained to mutual benefit in adverse weather. The iterative pseudo-label scheme is simple and appears effective across three datasets, and the architectural components (DDM, SAG, MFT, WPN) are clearly motivated. The Weather-Apollo dataset is a useful addition, and the FSS metric could be of interest to the community. However, the evidence base is narrow: only three baselines, no error bars, and the only real-world dataset relies on approximate cross-season training pairs. The FSS metric is also partially circular because it uses the same LPR model that provides pseudo-labels to the LDR. These issues prevent the state-of-the-art claim from being fully established as presented.","major_comments":[{"comment":"The LDR training pairs on Boreas are generated by pose-matching degraded Seq 02/03 scans to the nearest Seq 00 scan under strict pose thresholds (0.01 m, 0.1 deg), with the paper acknowledging that environmental changes may cause misalignment. Because Boreas is the only real-world dataset, the validity of these pairs is load-bearing for the real-world claim. The defense that the approximation is 'sufficient' is circular, since it cites the improved LPR numbers that were produced by the same pairs. Please report the number of matched pairs, show qualitative examples of the pairs, and provide a control experiment (e.g., an LDR trained only on the synthetic datasets and applied to Boreas) to isolate whether the real-world gains depend on valid cross-season pairs.","section":"§IV-A, Eq. (26)"},{"comment":"All results are single-run point estimates, and the Union-vs-Separate gains on Boreas are small (e.g., R@1 0.71 vs 0.65 easy, 0.63 vs 0.58 hard). Without multiple seeds or statistical significance measures, the claim that Union consistently outperforms Separate is not statistically supported. Please report means and standard deviations over at least three runs, or justify why single-run comparisons are sufficient in this setting.","section":"Tables II–IV"},{"comment":"The 'Separate' setting is described as pairing each LPR model with ITDNet-D 'via separate training,' but the paper never states whether this ITDNet-D is trained with the task-driven loss (Eq. 16) or only with the reconstruction loss (Eq. 15), nor the epoch budget for the separate modules. If the same jointly trained ITDNet-D is reused, the Union-vs-Separate comparison conflates joint optimization with the effect of the task-driven loss and is unfair. Please specify the exact training protocol for the Separate setting and retrain if necessary.","section":"§IV-C, Tables II–IV"},{"comment":"The FSS metric uses global descriptors from the same LPR model that provides pseudo-labels to the LDR during training, so the high FSS value for ITDNet may reflect feature-space overfitting rather than restoration quality. The paper's claim that FSS 'correlates well with downstream LPR performance' is based on only four methods and is therefore not established. The independent R@1 and R@1% columns in Table V are more reliable; please either remove the strong FSS claim or validate FSS against an independently trained descriptor (e.g., a frozen LPR not used in training).","section":"Eq. (27) and Table V"},{"comment":"The 'state-of-the-art' claim is supported by only three baselines (OT, LCDNet, CVTNet). Given the breadth of LiDAR place recognition literature, please add at least one or two standard recent methods (e.g., PointNetVLAD, MinkLoc3D, or a recent transformer-based descriptor) to the Direct/Separate comparisons, or temper the SOTA claim to 'superior to the compared methods.'","section":"§IV-C and Abstract"}],"minor_comments":[{"comment":"'We conduct qualitative evaluations on three large-scale LiDAR datasets' should be 'quantitative evaluations,' since the section reports tables of numeric metrics.","section":"§IV-A, first paragraph"},{"comment":"The notation p_j and \\hat p_j is confusing because p is used elsewhere for point clouds; these are feature vectors, so please rename them to clarify the distinction.","section":"Eqs. (16)–(17)"},{"comment":"Please specify whether epoch numbering starts at 1 so the odd/even parity is unambiguous; the text and algorithm should agree on which epoch trains which module.","section":"Algorithm 1"},{"comment":"The checkmark columns are not aligned with the component names, making it hard to see which component is removed in each row, and the text omits discussion of the MFT and WPN ablations.","section":"Table VI"},{"comment":"The caption says 'the restored image is decomposed,' but the LDR operates on range images, not natural images; please rephrase to 'the restored range image features.'","section":"Figure 2(b) caption"},{"comment":"The GitHub URL is broken across lines; please provide it as a single URL. Also, if a peer-reviewed version of reference [7] (ResLPRNet) exists, please cite it in addition to the arXiv preprint.","section":"Abstract and references"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid systems paper, but the evidence base needs strengthening. The Boreas pairing concern is the most serious; I would like to see either evidence that the pose-matched pairs are semantically valid or a synthetic-only control experiment. The missing error bars are also important given the small margins on Boreas. The FSS metric is a minor contribution and could be downplayed. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new contribution is the iterative joint training of restoration and recognition: the LDR module cleans the range image, the LPR module feeds feature pseudo-labels back to guide restoration, and the two alternate. That combination is not in ResLPRNet or the other cited works, and it works. The central ordering—Union beats Separate beats Direct—holds in essentially every configuration across Weather-KITTI, Weather-Apollo, and Boreas, with dramatic jumps like Seq 00 snow R@1 from 0.15 to 0.84. The frequency-domain blocks are new assemblies of known ingredients, and the ablations show each one earns its keep.\n\nSoft spots, in proportion. No error bars and only three LPR baselines, so the strength rests on consistency across datasets rather than statistical rigor—but the consistency is there. The Boreas LDR pairs are approximate (closest clean scan in another season within 0.01 m and 0.1°), and the paper's defense that the improved Table IV numbers prove sufficiency is a little circular. However, the retrieval metrics are on an unseen test sequence, so the loop isn't closed in a damaging way. Still, a reviewer should want a few manual checks of those pairs or a sensitivity analysis. The FSS metric is self-referential: it measures feature similarity with the same LPR descriptors being trained, so Table V's restoration comparison is partly circular. The R@1 column is independent and carries the real evidence. Code and data are promised but not yet released, so the results are not independently reproducible today.\n\nThe math is standard, the derivations are transparent, and no result reduces to a fitted parameter. This is a solid engineering contribution for anyone working on LiDAR perception in bad weather. I'd send it to peer review; the soft spots are fixable.","headline":"Iterative joint training of restoration and recognition is a real, well-supported contribution; the Boreas pairing and self-referential FSS are fixable soft spots.","tokens_in":19374,"tokens_out":3510,"would_cite":true,"duration_ms":32551,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that joint, alternating training of LiDAR restoration and place recognition beats separate pipelines in rain, snow, and fog.","keywords":["LiDAR place recognition","adverse weather","LiDAR data restoration","iterative task-driven learning","task-driven loss","frequency-domain feature mixing","wavelet pyramid descriptor","range image"],"falsifier":"Remove the task-driven KL loss from the LDR training on Boreas while keeping everything else fixed; if recall on the hard split does not fall below the reported 0.63 R@1, the iterative coupling is not what produces the Union gain.","tokens_in":18397,"feed_emoji":"🌨️","tokens_out":8968,"duration_ms":76186,"temperature":0.7,"pith_summary":"The paper tries to establish that LiDAR place recognition in rain, snow, and fog fails partly because restoration and recognition are trained as separate stages, so the restored range image is not optimized for what the recognizer needs. Its framework, ITDNet, alternates between training a range-image restoration module and a place-recognition descriptor module in successive epochs, letting the recognizer's global descriptors act as pseudo-labels that supervise the restorer through a KL-divergence task-driven loss. On synthetic Weather-KITTI and Weather-Apollo and on real-world Boreas, this joint \"Union\" mode consistently beats both direct recognition on degraded scans and a separate restore-then-recognize pipeline; for example, snow R@1 on Weather-KITTI Sequence 00 rises from 0.15 direct and 0.74 separate to 0.84 Union. If the claim holds, adverse-weather localization is a task-coupling problem, not just a denoising or descriptor problem.","feed_headline":"Coupling LiDAR cleanup and place recognition lifts bad-weather recall","feed_subtitle":"On snowy KITTI, top-1 recall jumps from 0.74 separate to 0.84 when both modules train together.","key_machinery":"The load-bearing mechanism is the alternating optimization loop between the LDR and LPR modules. The LDR module is a U-Net built from Dual-Domain Mixer (DDM) blocks, which alternate FFT-based frequency mixing with depthwise spatial mixing to suppress high-frequency weather noise, and Semantic-Aware Generator (SAG) blocks, which inject multi-scale semantic context. The LPR module applies two-level wavelet decomposition, refines each sub-band with Multi-Frequency Transformer (MFT) blocks using frequency-guided window attention, and aggregates the scales through a Wavelet Pyramid NetVLAD (WPN) block. The two are tied together by the task-driven loss, a KL divergence between the softmaxed global descriptors of restored and clean scans produced by the frozen LPR teacher, combined with an L1 reconstruction loss on range and intensity, with the task-driven weight raised from 0.01 to 0.1 after the first 30 epochs.","core_discovery":"The central claim is that LiDAR data restoration and LiDAR place recognition should be optimized together, in alternating epochs, with each task shaping the other. In even epochs the LPR module is trained with triplet loss on restored queries and clean database scans; in odd epochs the LDR module is trained with a reconstruction loss plus a task-driven loss that pulls the global descriptor of the restored scan toward the descriptor of the clean scan, using the previous epoch's LPR module as a frozen teacher. The paper reports that this Union mode outperforms Direct use of degraded scans and Separate restoration-then-recognition on every tested dataset and weather condition, and it introduces the Feature Similarity Score to show that alignment in feature space, rather than pixel fidelity, tracks recognition gains.","pith_inferences":["The same alternating pseudo-label coupling could be carried over to other degraded-input perception pairs, such as dehazing with object detection or desnowing with semantic segmentation; the paper surveys union-learning precedents but does not test those pairs.","Using the Feature Similarity Score as a training reward could let the restorer learn without clean paired scans, which would directly address the cross-season pairing problem the paper acknowledges on the real-world dataset.","Because the Union gains over Separate are larger on synthetic data than on Boreas, the practical ceiling may sit in how well degraded and clean scans can be paired; improving pair generation or moving to self-supervised restoration is a plausible next step.","The task-driven loss may be acting mainly as a semantic regularizer on the restorer; if so, a lighter feature-matching penalty could capture most of the Union gain at lower training cost."],"forward_implications":["A navigation stack using ITDNet could keep recognizing places in snow, fog, and rain at substantially higher top-1 recall than running restoration and recognition as separate stages.","Restoration quality for place recognition should be measured by feature-level alignment rather than pixel-level similarity; the paper's Feature Similarity Score predicts downstream retrieval better than SSIM.","The Union advantage appears across synthetic and real-world data and across snow, fog, and rain, so the benefit is not tied to a single corruption type.","Range-image-based recognizers gain more from the restoration module than point-based recognizers do, pointing to image-based pipelines as the natural integration target.","At roughly 24 ms per scan with about 29.86 million parameters, the full framework is fast enough for real-time onboard use."],"supporting_citations":[{"why":"Cross-view transformer baseline CVTNet that ITDNet is compared against in Direct, Separate, and Union settings.","marker":"[1]"},{"why":"ResLPRNet, the prior separate restoration-for-place-recognition network ITDNet's joint training is designed to supersede.","marker":"[7]"},{"why":"4DenoiseNet, a LiDAR denoising baseline in the restoration comparison.","marker":"[12]"},{"why":"TripleMixer, the source of the Weather-KITTI degradation simulation and a denoising baseline.","marker":"[13]"},{"why":"NetVLAD, the descriptor aggregation method used inside the WPN block.","marker":"[24]"},{"why":"OverlapTransformer, a range-image LPR baseline used in all three evaluation modes.","marker":"[26]"},{"why":"LCDNet, a point-based LPR baseline used in all three evaluation modes.","marker":"[45]"},{"why":"Boreas, the real-world multi-season dataset for the main real-world evaluation.","marker":"[46]"},{"why":"SemanticKITTI, source of the clean scans paired with Weather-KITTI degraded sequences.","marker":"[47]"},{"why":"SSIM, the pixel-level metric the paper contrasts with its Feature Similarity Score.","marker":"[51]"}],"fun_headline_variants":["Snowy KITTI recall jumps from 0.74 to 0.84 with joint training","Joint LiDAR restoration and recognition beat separate pipelines in bad weather","Alternating epochs tie LiDAR restoration to place recognition gains","Coupling cleanup and recognition lifts bad-weather place recall","Iterative task loop makes LiDAR place recognition robust in snow and fog"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world claim rests on the assumption that pairing each degraded Boreas scan with the nearest clean scan from another season (within 0.01 m and 0.1 degrees) yields genuinely matching scene pairs, an assumption the paper itself flags as vulnerable to environmental misalignment.","fun_headline_variants_meta":{"raw":{"variants":["Snowy KITTI recall jumps from 0.74 to 0.84 with joint training","Joint LiDAR restoration and recognition beat separate pipelines in bad weather","Alternating epochs tie LiDAR restoration to place recognition gains","Coupling cleanup and recognition lifts bad-weather place recall","Iterative task loop makes LiDAR place recognition robust in snow and fog"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00117,"raw_usage":{"total_tokens":4875,"prompt_tokens":1017,"completion_tokens":3858,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":3764}},"tokens_in":633,"tokens_out":3858,"duration_ms":27946,"temperature":1.0,"reasoning_tokens":3764,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:40.020475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove the task-driven KL loss from the LDR training on Boreas while keeping everything else fixed; if recall on the hard split does not fall below the reported 0.63 R@1, the iterative coupling is not what produces the Union gain.","supporting_citations":[{"cited_title":"Cvtnet: A cross-view transformer network for lidar-based place recognition in autonomous driving envi- ronments,","cited_arxiv_id":null,"evidence_quote":"Cross-view transformer baseline CVTNet that ITDNet is compared against in Direct, Separate, and Union settings."},{"cited_title":"ResLPR: A LiDAR Data Restoration Network and Benchmark for Robust Place Recognition Against Weather Corruptions","cited_arxiv_id":"2503.12350","evidence_quote":"ResLPRNet, the prior separate restoration-for-place-recognition network ITDNet's joint training is designed to supersede."},{"cited_title":"4denoisenet: Adverse weather denoising from adjacent point clouds,","cited_arxiv_id":null,"evidence_quote":"4DenoiseNet, a LiDAR denoising baseline in the restoration comparison."},{"cited_title":"Overlaptransformer: An efficient and yaw-angle-invariant transformer network for lidar-based place recognition,","cited_arxiv_id":null,"evidence_quote":"OverlapTransformer, a range-image LPR baseline used in all three evaluation modes."},{"cited_title":"Lcdnet: Deep loop closure detection and point cloud registration for lidar slam,","cited_arxiv_id":null,"evidence_quote":"LCDNet, a point-based LPR baseline used in all three evaluation modes."},{"cited_title":"Boreas: A multi-season autonomous driving dataset,","cited_arxiv_id":null,"evidence_quote":"Boreas, the real-world multi-season dataset for the main real-world evaluation."}],"review_version":1}