{"id":"67ab41e8-b697-447c-a2f1-6123a17826c7","arxiv_id":"2412.01299","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LiDAR intensity textures, projected with HEC and matched by pretrained visual features, achieve cross-modal relocalization in prior LiDAR maps with high recall on two self-collected campus datasets.","lead":"This paper builds a system that lets a single camera image find its position inside a pre-built LiDAR map, by turning the map's intensity information into panorama-like images and matching them to the camera picture. It could help robots and autonomous vehicles re-localize using only a cheap camera when a detailed laser map already exists.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth camera poses are derived from LiDAR localization (Sec. IV-A) but no camera-LiDAR extrinsic or query-time LiDAR scan is reported; all RR and R@K metrics depend on this transform, so the quantitative evidence for the central claim is not verifiable as written.","rationale":"The reader's weakest assumption is the zero-shot intensity-grayscale transfer and its generalization beyond two campus sequences. That is an external-validity concern. The concern I raise is internal: even on the reported datasets, the ground-truth camera poses may not actually be camera poses because the paper never specifies the camera-to-LiDAR extrinsic or the query-time LiDAR acquisition. If the ground truth is the LiDAR pose, all reported R@K and RR values are biased, and the central claim—effective visual relocalization—is not demonstrated. This is more load-bearing than the generalization issue because it affects the validity of every quantitative result, not just their breadth. The method itself is presented clearly and the intensity-versus-depth ablation is a reasonable control, which is why I do not recommend rejection. The verdict remains CONDITIONAL: the authors should disclose the extrinsic calibration and query sensor setup, ideally by releasing data or evaluating on a public benchmark with known extrinsics. This additional condition is consistent with the reader's CONDITIONAL verdict, but for a different reason.","tokens_in":9496,"tokens_out":14236,"duration_ms":124548,"concrete_test":"Request the camera-to-LiDAR extrinsic matrix and the query-time LiDAR poses used to generate the ground truth in Section IV-A. Recompute Table I and Table III by applying that extrinsic to the LiDAR poses. If the recomputed R@1 or RR(1m/3°) differ from the published values by more than the respective thresholds (5 m or 1 m/3°), the evaluation is not measuring camera relocalization and the central claim is unsupported. If the extrinsic is already accounted for and the numbers are unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states that 'our dataset was annotated with ground truth camera poses, which were obtained by map-based LiDAR localization within the same prior maps.' This is the sole source of ground truth for every reported metric: Table I defines retrieval correctness using a 5 m translation error, and Table III uses RR thresholds of 0.5 m/1° and 1 m/3°. However, the paper never reports (i) whether a LiDAR scan was captured alongside the query camera images, (ii) the extrinsic calibration between the camera and the LiDAR, or (iii) the expected accuracy of the LiDAR localization itself. Without the extrinsic transform, the 'camera poses' may in fact be LiDAR poses, which would introduce a constant offset and rotation into every error computation. Given the tight thresholds, even an unstated 0.3 m / 1° offset could materially change RR. This is a missing support in the evidence for the central claim; the paper's own Section IV-A omits it, preventing verification and reproduction. The intensity-versus-depth comparison is a useful internal control, but the absolute values of the headline metrics remain unsubstantiated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical cross-modal visual relocalization system that estimates a camera's 6DoF pose inside a pre-built LiDAR map. The map is first rendered as panoramic intensity images using HEC projection, then coarse candidate retrieval is performed with NetVLAD and a covisibility-clustering reranking step, and fine pose estimation is obtained from SuperPoint+LightGlue two-stage 2D-3D association followed by covisibility inlier selection and PnP with RANSAC. Experiments are conducted on two self-collected SJTU campus sequences, reporting R@1 of 0.73 on the College sequence and relocalization recall RR(1m/3°) of 0.97 on College and 0.79 on Dormitory, along with an ablation on the College sequence.","tokens_in":9751,"tokens_out":5927,"duration_ms":56958,"significance":"If the reported results hold, the paper would show that pretrained visual feature extractors and matchers can be applied to LiDAR intensity images without any fine-tuning, which is practically valuable for cross-modal relocalization. The HEC projection and two-stage equalization are reasonable engineering contributions aimed at reducing the modality gap, and the intensity-versus-depth internal comparison in Table III is a useful control that isolates the role of intensity texture. The main significance is limited by the evaluation being restricted to two private sequences with no public benchmark, no error bars, and no independent ground-truth verification, so the generality of the claims is not yet established.","major_comments":[{"comment":"Every quantitative result in Tables I and III depends on the ground-truth camera poses, which Section IV-A describes only as 'obtained by map-based LiDAR localization within the same prior maps.' The paper does not report whether a LiDAR scan was captured together with each query camera image, the extrinsic calibration between the camera and the LiDAR, or the expected accuracy of the LiDAR localization itself. If the quoted poses are LiDAR poses rather than camera poses, then a constant camera-to-LiDAR offset and rotation enter all error computations; with RR thresholds as tight as 0.5 m and 1 degree, even a small unmodeled offset could materially change the reported RR values. The authors should specify the sensor setup, report the extrinsic calibration and its residual error, and validate the ground-truth accuracy, for example by showing reprojection errors against known targets or by providing the LiDAR localization residuals.","section":"IV-A"},{"comment":"The quantitative evaluation covers only two self-collected sequences on a single campus, and no error bars, query counts, or per-query error distributions are reported. The ablation in Table II is performed on only the College sequence. As a result, the paper's broader claim that intensity-texture consistency enables reliable relocalization is not supported beyond these two environments. The authors should report the number of queries, give confidence intervals or repeated-run statistics, and ideally evaluate on at least one independent or public dataset, or on additional sequences with different point-cloud density and appearance conditions.","section":"IV-B"},{"comment":"The fine-relocalization comparison is limited to a depth-channel variant of the authors' own pipeline, which is an internal control rather than a state-of-the-art cross-modal localization baseline. The introduction and conclusion claim superiority over existing methods, but Tables I and III do not establish this for pose estimation because no existing cross-modal relocalization or 2D-3D registration method is compared at the pose level. The authors should either add comparisons to published cross-modal localization/registration baselines or explicitly restrict the claim to 'the intensity channel improves over the depth channel within this hierarchical pipeline.'","section":"IV-B, Table III"},{"comment":"The ablation study shows small differences for several components: removing covisibility clustering and removing two-stage equalization both lower RR(1m/3°) from 0.97 to 0.92 on a single sequence. Without error bars, repeated trials, or additional sequences, these differences are not clearly distinguishable from noise. At the same time, the 0.5m/1° RR values show large drops for some ablations (e.g., from 0.25 to 0.06 without two-stage 2D-3D association), which suggests the component has a strong effect only at very tight thresholds. The authors should provide statistical support and discuss which conclusions are robust to the choice of threshold.","section":"Table II"}],"minor_comments":[{"comment":"The phrase 'Struct-from-Motion' should be 'Structure-from-Motion,' and the sentence 'we propose a hierarchical ... system' would benefit from grammatical cleanup.","section":"I"},{"comment":"'Habrid Equiangular Cube Projection' appears to be a typo for 'Hybrid Equiangular Cube Projection'; also, the variables in Equations (2) and (3) and the meaning of the adjusting parameter t should be defined precisely.","section":"III-B"},{"comment":"The 'sliding window strategy' used to maximize similarity scores for four cube patches is described only in one sentence; the window size, stride, and how the four scores are combined should be stated.","section":"III-C1"},{"comment":"The criteria for discarding a query-map pair when 'the largest cluster does not contain enough matches' are not specified; the threshold should be reported.","section":"III-D1"},{"comment":"Covisibility of a local feature point is defined as the number of map images that contain the point, but the paper does not explain how this count is computed for points on continuous surfaces or how the covisibility threshold is chosen.","section":"III-D2"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper's core pipeline is sound and built on well-established components, so the main issue is not technical correctness but the strength of the empirical evidence. The missing extrinsic calibration and ground-truth accuracy documentation is a blocker for the quantitative claims, and the baselines need to be framed more honestly. If the authors provide the missing sensor/calibration details and strengthen the evaluation with error bars and additional data, the paper could become acceptable. I do not see grounds for rejection, since the internal intensity-versus-depth comparison is useful and the proposed system demonstrates a plausible practical approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: this is a practical systems paper, not a methodological breakthrough. The new bit is treating LiDAR intensity as a texture channel, projecting it with HEC into panoramic images, and then running off-the-shelf NetVLAD/SuperPoint/LightGlue on those images to relocalize a single camera image in a prior LiDAR map. The components are all known, but the composition is sensible, and the intensity-versus-depth comparison shows the intensity channel does carry the load. On the paper's own two campus sequences, the headline numbers (R@1 0.73, RR(1m/3°) 0.97 on College) are plausible.\n\nWhat the paper does well: it is honest about using pretrained models with no fine-tuning, which is a strong transfer claim, and it actually tests that claim by comparing against a depth-channel variant of the same pipeline. The covisibility clustering and two-stage matching are reasonable engineering choices, and the ablation on College isolates their contributions. The writing is clear and the related work is properly situated.\n\nWhere it gets soft: the evaluation is thin. Two self-collected sequences, no public code or data, no error bars, and the ablation is on one sequence. The pose baseline is a depth-channel version of the same pipeline, not an existing cross-modal relocalization method, so the 'we beat prior art' story is really 'we beat a depth baseline.' And the stress-test note is on point: ground truth camera poses come from map-based LiDAR localization, but the paper never says whether a LiDAR scan was captured with each query image, what the camera-LiDAR extrinsic was, or how accurate that localization is. With RR thresholds as tight as 0.5m/1°, an unstated offset could shift the numbers. That's a verifiability gap, not a demonstrated error, but it matters.\n\nThe central idea—that intensity textures bridge the grayscale-image/LiDAR-map gap—is credible and probably useful for SLAM initialization in prior maps. The paper would benefit from a public benchmark, more sequences, error bars, and a real prior-art pose-estimation baseline. As is, it's a solid workshop-to-conference paper, not a definitive result.\n\nI'd send it to review—it's coherent, reproducible in principle, and the intensity-texture idea deserves more attention. But I'd ask the authors to release data/code or add a public-benchmark evaluation, and to document the ground-truth pipeline before the numbers can be taken at face value.","headline":"A sensible intensity-texture pipeline for camera-to-LiDAR-map relocalization, but the evidence is narrow and the ground-truth provenance is under-reported.","tokens_in":10332,"tokens_out":2070,"would_cite":false,"duration_ms":17998,"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":"LiDAR intensity textures let a camera relocalize in prior maps","keywords":["visual relocalization","LiDAR intensity","cross-modal localization","place recognition","pose estimation","HEC projection","2D-3D association","6DoF pose"],"falsifier":"Run the exact pipeline with a sparser or differently structured LiDAR map (for example, a 16-beam Velodyne scan of a highway or tunnel) and measure RR(1m/3°) on a comparably sized query set; if the recall drops below roughly 0.5, the intensity-texture transfer claim collapses. Alternatively, replace HEC with equirectangular projection while keeping every other module identical and check whether the recall drop matches the paper's ablation, where RR(1m/3°) falls from 0.97 to 0.61 on the College sequence.","tokens_in":60,"feed_emoji":"🗺️","tokens_out":3231,"duration_ms":82945,"temperature":0.7,"pith_summary":"The paper tries to establish that a single camera image can be globally relocalized inside a pre-built LiDAR map by exploiting the intensity channel of the point cloud as a texture, without retraining any visual network. It argues that when the map is projected through HEC panoramic projection and equalized in two stages, the resulting intensity images become similar enough to grayscale camera images that off-the-shelf retrieval and matching networks work reliably. If true, this gives a complete hierarchical pipeline: project the map into intensity images, retrieve candidate places, match 2D features to 3D points, and estimate the 6DoF pose. The reported results, with 73% top-1 retrieval on the College sequence and 97% relocalization recall at 1m/3°, support the claim on the paper's own datasets.","feed_headline":"LiDAR intensity textures let a camera relocalize in prior maps","feed_subtitle":"Off-the-shelf visual networks retrieve and match grayscale-style intensity images, hitting 97% pose recall on campus scenes.","key_machinery":"The central object is the intensity-channel panoramic map image produced by HEC projection from the prior LiDAR map. HEC adjusts cube-face sampling with an arctan-based formula to densify and homogenize the projected pixels, discarding the top and ground surfaces, and the two-stage equalization then aligns the intensity histogram with grayscale image appearance. This projected image acts as the proxy that brings camera pixels and LiDAR points into one modality, allowing standard visual place recognition and matching networks to carry cross-modal association without special cross-modal training.","core_discovery":"The central claim is that intensity-texture consistency between LiDAR intensity images and grayscale camera images is sufficient for cross-modal visual relocalization, provided the map is projected with HEC (Hybrid Equiangular Cube Projection) and equalized in two stages: a 3D map histogram equalization for initial normalization followed by CLAHE for contrast enhancement. This makes pretrained visual models directly applicable: NetVLAD for coarse retrieval, SuperPoint for local feature extraction, and LightGlue for matching, with the resulting 2D-2D correspondences lifted to 2D-3D using the stored map points. The paper reports Recall@1 of 0.73 on the College sequence and relocalization recall RR(1m/3°) of 0.97 and 0.79 on the two campus sequences, clearly exceeding a depth-channel variant of the same pipeline.","pith_inferences":["The reliance on pretrained visual networks implies the strongest test is LiDAR sparsity: on a 16-beam or otherwise sparser sensor, the intensity images would have larger holes and the texture consistency would degrade, likely dropping both retrieval and pose recall.","The paper's own stated future work of retraining retrieval and matching networks to learn intensity-grayscale consistency suggests the authors see the current off-the-shelf transfer as effective but not optimal; a specialized network could push recall higher on harder scenes.","The covisibility inlier selection idea, which prunes matches by how many map images observe a point, is not tied to intensity specifically and could generalize to any cross-modal matching task where a texture proxy image is built from a 3D map."],"forward_implications":["If the intensity-texture transfer holds, existing pretrained visual localization stacks can be reused for LiDAR-map relocalization with zero fine-tuning.","The intensity channel can replace depth or geometry as the primary cross-modal bridge, simplifying the 2D-3D association step.","The complete hierarchical pipeline offers single-frame global localization, not just place recognition or registration alone, which is directly useful for loop closure and global pose initialization in LiDAR SLAM.","The method's success on campus scenes suggests that similar intensity-based texture proxies could be adopted in other map-based localization settings where camera images must be matched against non-visual sensors."],"supporting_citations":[{"why":"Provides the NetVLAD global descriptor used for coarse retrieval of map images against the query image.","marker":"[24]"},{"why":"Provides the SuperPoint local feature extractor used to detect keypoints in both query and map images.","marker":"[15]"},{"why":"Provides the LightGlue matcher used in the two-stage 2D-2D association between query and map images.","marker":"[16]"},{"why":"Defines the HEC projection model that the paper adopts to produce denser, less distorted panoramic map images.","marker":"[23]"},{"why":"Supplies the LIP-loc cross-modal baseline whose depth-channel map images show poor generalization in the experiments.","marker":"[8]"},{"why":"Supplies the Patch-NetVLAD retrieval baseline that the proposed method outperforms on both campus datasets.","marker":"[25]"}],"fun_headline_variants":["Intensity textures bridge camera and LiDAR for relocalization","Match LiDAR intensity maps to camera images for relocalization","Camera relocalization using LiDAR intensity textures","HEC-projected intensity maps enable cross-modal visual relocalization","Intensity equalization makes LiDAR maps camera-ready for relocalization"],"cache_read_input_tokens":12416,"weakest_assumption_plain":"The entire pipeline assumes that after HEC projection and two-stage equalization, LiDAR intensity images are texture-consistent enough with grayscale camera images that off-the-shelf visual networks can match them reliably, and this assumption is tested only on two self-collected campus sequences using a Livox Avia LiDAR.","fun_headline_variants_meta":{"raw":{"variants":["Intensity textures bridge camera and LiDAR for relocalization","Match LiDAR intensity maps to camera images for relocalization","Camera relocalization using LiDAR intensity textures","HEC-projected intensity maps enable cross-modal visual relocalization","Intensity equalization makes LiDAR maps camera-ready for relocalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000606,"raw_usage":{"total_tokens":2810,"prompt_tokens":914,"completion_tokens":1896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1810}},"tokens_in":530,"tokens_out":1896,"duration_ms":12697,"temperature":1.0,"reasoning_tokens":1810,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:30:09.924510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact pipeline with a sparser or differently structured LiDAR map (for example, a 16-beam Velodyne scan of a highway or tunnel) and measure RR(1m/3°) on a comparably sized query set; if the recall drops below roughly 0.5, the intensity-texture transfer claim collapses. Alternatively, replace HEC with equirectangular projection while keeping every other module identical and check whether the recall drop matches the paper's ablation, where RR(1m/3°) falls from 0.97 to 0.61 on the College sequence.","supporting_citations":[{"cited_title":"Superpoint: Self- supervised interest point detection and description,","cited_arxiv_id":null,"evidence_quote":"Provides the SuperPoint local feature extractor used to detect keypoints in both query and map images."},{"cited_title":"Efficient projection and coding tools for 360° video,","cited_arxiv_id":null,"evidence_quote":"Defines the HEC projection model that the paper adopts to produce denser, less distorted panoramic map images."}],"review_version":1}