{"id":"49c4cc80-001e-4f9f-8abd-37587f17e85a","arxiv_id":"2505.03300","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A pipeline segments aligned LiDAR point clouds by rendering intensity-colored 2D views, applying a camera-domain 2D segmentation model, and voting back-projected labels, producing competitive pseudo-labels for unsupervised domain adaptation.","lead":"Researchers generate 2D images from LiDAR scans colored by sensor intensity, segment them with a standard 2D model trained on camera images, and project the labels back to 3D to create pseudo-labels without any 3D annotation. The method is demonstrated on nuScenes and in a SemanticKITTI-to-nuScenes domain adaptation setting, reaching mIoU 61.4 when pretrained.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Domain transfer from greyscale camera images to LiDAR intensity views is never measured in isolation; final 3D metrics confound it with multi-view voting.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the 2D-to-LiDAR domain transfer is assumed but not isolated. My analysis agrees. The paper's empirical results (Fig. 2, Table I, Table II) provide some end-to-end evidence, but they cannot distinguish the contribution of the domain-transfer assumption from that of the multi-view voting and the favorable evaluation conditions. The manuscript itself flags the domain-gap issue in the conclusion, which underscores that this is a known limitation. A per-view evaluation is the direct test: it would either validate the core premise or show that the method relies on voting and cropping rather than on genuine 2D recognition of intensity views. Other concerns raised by the reader, such as the weak Naive baseline and missing error bars, are secondary because they affect the strength of the UDA comparison rather than the foundational assumption. The verdict should remain CONDITIONAL, as the paper is a plausible proof-of-concept but the central assumption is not yet established.","tokens_in":10042,"tokens_out":3278,"duration_ms":31652,"concrete_test":"Render the K=600 virtual views for a random subset of 10 nuScenes scenes using the paper's exact pose generation settings, project the 3D ground-truth labels into each 2D view to create pixel-level ground truth, and run the trained Mask2Former 2D model on these rendered intensity views to compute per-view mIoU and per-class IoU before any voting or back-projection. Compare these numbers to the same model's mIoU on a held-out validation set of greyscale-converted Mapillary Vistas images. If the per-view mIoU on LiDAR intensity views is within 10 points of the camera-domain mIoU, the domain-transfer assumption holds; if it is substantially lower, the final 3D metrics are primarily due to the election estimator and the restricted evaluation, weakening the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the assumption that a 2D model trained on greyscale-converted Mapillary Vistas images will segment rendered LiDAR intensity views correctly. This assumption enters in Section III and Section IV-A1, where the authors state they 'convert the RGB images to greyscale images to match the LiDAR intensity domain' and 'believe this diversity can help bridge the domain gap between camera images and generated views.' The belief is never tested directly. All reported quality metrics in Fig. 2 and Table I are computed after multi-view voting and 3D back-projection, which confounds the 2D model's per-view accuracy with the fusion mechanism and the evaluation setup (30 m crop, merged sidewalk/terrain, static classes only). If the 2D model performs poorly on individual intensity views, voting can still produce acceptable 3D results for easy classes like road, but the pipeline's claim of general '3D semantic segmentation' would be overstated. The ablation in Table I shows a 21% mIoU drop when augmentations are removed, suggesting the gap is significant but not quantified. The conclusion itself admits 'the limited domain in which the views can be generated without increasing the domain shift with the 2D dataset,' which is a direct acknowledgment that the domain-transfer constraint is unresolved. Without a separate measurement of the 2D model's accuracy on rendered views, the core premise of the method remains unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for generating pseudo-labels for LiDAR point cloud semantic segmentation without 3D annotations or camera images at inference time. Point clouds from a sequence of scans are aligned and colored only by sensor intensity; many virtual views are rendered from randomized poses; a 2D segmentation model trained on greyscale-converted Mapillary Vistas images is applied to each view; the resulting labels and logits are back-projected and fused by voting. The pseudo-labels are evaluated against nuScenes ground truth on five static classes, and their usefulness is tested in an unsupervised domain adaptation setting (SemanticKITTI to nuScenes), where the method reaches 61.4 mIoU with pretraining, comparable to T-UDA.","tokens_in":10326,"tokens_out":4195,"duration_ms":42166,"significance":"If the central premise holds, the method is a meaningful contribution: it is the first, to the authors' knowledge, to propagate 2D labels to generated LiDAR-intensity views, removing the need for posed camera images at inference. The paper provides a useful ablation study, evaluates on a large dataset, and is honest about limitations. However, the load-bearing domain-transfer assumption is never measured in isolation, the experiments are single-run without error bars, and the comparison to T-UDA is complicated by a large naive-baseline discrepancy. These issues make the quantitative claims less definitive than the text suggests.","major_comments":[{"comment":"The domain-transfer assumption that a 2D model trained on greyscale-converted Mapillary Vistas images can segment rendered LiDAR-intensity views is never measured directly. Every reported quality metric in Fig. 2 and Table I is computed after multi-view voting and 3D back-projection, so the per-view accuracy of the 2D model is confounded with the fusion estimator. The Conclusion's own admission that 'the limited domain in which the views can be generated without increasing the domain shift with the 2D dataset' remains an unresolved challenge is itself evidence that the key premise needs separate validation. Please add a direct evaluation of the 2D model on individual rendered views before fusion, for example per-view IoU on a held-out subset of generated views, and report the numbers per class.","section":"Section III / IV-A1 / Conclusion"},{"comment":"All ablation results come from a single run with no error bars, multiple seeds, or significance tests. Several pairwise differences in Table I, such as (d) at 49.95 mIoU versus (e) at 49.77, and (f) at 53.14 versus (g) at 53.12, are small enough to be plausibly within run-to-run variance, yet the text draws conclusions from them, including the preference for 'summing soft votes' over 'compounding soft votes'. Please provide confidence intervals or repeated-seed results and base the claims about estimator choice and randomization levels on those results.","section":"Table I / Section IV-A2"},{"comment":"The comparison with T-UDA [6] is not apples-to-apples because the authors report a Naive no-DA mIoU of 7.0 while [6] reports 34.18 on the same classes, a discrepancy the paper attributes to pretraining augmentations without evidence. If the source-pretrained model used here is not trained under the same protocol as in [6], the UDA improvements in Table II are not directly comparable to the published numbers. Please reproduce the pretraining protocol of [6], or report the [6] Naive baseline alongside the authors' own Naive baseline in the same table, and temper the statement that the method is 'on-par' with T-UDA accordingly.","section":"Table II / Section IV-B2"},{"comment":"The headline pseudo-label mIoU is computed on a cropped point set (points beyond 30 m on each side and 10 m in height are removed) and after merging sidewalk and terrain, while the abstract and introduction claim '3D semantic segmentation' of whole scenes. These evaluation choices are post-hoc and are not reflected in the stated mIoU. Please report uncropped and unmerged metrics, or explicitly state the restricted evaluation domain in the abstract and conclusion so readers can judge the generality of the claim.","section":"Fig. 2 / Section IV-A / Abstract"}],"minor_comments":[{"comment":"The figure reports an average IoU of 0.78 for the road class, while the text states an average IoU of 0.81 for the same class; please harmonize the numbers.","section":"Fig. 2 / Section IV-A2"},{"comment":"The pose-sampling algorithm uses the interval '1 < i < M', which should presumably be '1 <= i <= M', and the number of sampled poses K is not otherwise constrained by M; please clarify the intended range.","section":"Algorithm 1"},{"comment":"The header row of Table I is ambiguous because the column labels are displayed together with check-mark rows rather than as clearly separated column names; please render the table with explicit column headers.","section":"Table I"},{"comment":"The '1h' baseline row in Table II is not defined in the text; please specify how many annotated scans it uses and how the model is trained.","section":"Table II"},{"comment":"The compounding estimator multiplies raw logits from each view, but no normalization is described before the product; the text mentions a possible missing prior normalization, but please state explicitly whether softmax-normalized probabilities or raw logits were compounded.","section":"Section III"},{"comment":"The abstract contains wording errors such as 'not other modality for inference'; please revise the language throughout for clarity and grammatical correctness.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real but narrow contribution, and it holds up well enough to send to review. The new thing is replacing camera RGB images with rendered LiDAR-intensity views for 2D-to-3D label propagation, which removes the camera requirement at inference. The experiments support the main claim, though one central assumption is never tested directly.\n\nThe method is simple: align a LiDAR sequence into a dense scene, color each point by clipped and normalized intensity, render 600 virtual camera views, run a Mask2Former trained on Mapillary Vistas (converted to greyscale), back-project labels and logits, and fuse by summing soft votes. The fusion itself is standard multi-view score fusion; the novelty is the input representation. The paper does a proper ablation over K, augmentations, pose noise, and fusion estimators, and it evaluates the pseudo-labels both directly against nuScenes ground truth and as training signal for UDA. That is the right test. The numbers (0.49 mIoU direct, 61.4 mIoU after UDA with SemanticKITTI pretraining) are plausible and competitive with T-UDA at 64.5. The evaluation against an independent dataset means the result is not circular.\n\nThe soft spots are clear. The load-bearing assumption is that a 2D model trained on greyscale camera images will segment rendered intensity views. This is never measured in isolation; every reported quality metric comes after multi-view voting and 3D back-projection, so per-view accuracy is unknown. The stress-test note is right about this. When augmentations are removed, mIoU drops 21%, which suggests the gap is real but not quantified. Second, there is one run, no error bars, and no code release. Third, the evaluation crops at 30m and merges sidewalk/terrain; this is disclosed, so it is a limitation rather than a trick, but it does make the numbers look better than an uncropped full-class evaluation would. Fourth, the Naive baseline discrepancy with T-UDA is hand-waved as 'difference in augmentations.' Since the UDA comparison is the main evidence of practical value, that explanation is too thin and needs more support.\n\nFor a reader: this paper is for people working on pseudo-label generation, UDA for LiDAR segmentation, or 2D-to-3D label transfer. It deserves a serious referee. I would ask for a per-view accuracy study, error bars across seeds, and a better explanation of the baseline mismatch. I would cite it if I were writing about zero-annotation LiDAR segmentation.","headline":"Real but narrow novelty—intensity-rendered views instead of camera images—with a solid empirical core and one untested domain-transfer assumption.","tokens_in":10883,"tokens_out":3005,"would_cite":true,"duration_ms":28091,"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 2D segmentation model trained on street images can label 3D LiDAR scenes when the point cloud is rendered as intensity-only views.","keywords":["LiDAR point clouds","3D semantic segmentation","pseudo-label generation","2D-to-3D label transfer","sensor intensity rendering","multi-view voting","unsupervised domain adaptation","driving scenes"],"falsifier":"Take the same aligned scenes and replace each point's real intensity with random values or a flat color, run the same 2D model and voting pipeline, and compare pseudo-label IoU to the reported roughly 0.49; if the score barely drops, the sensor intensity is not what drives the recognition, contradicting the paper's domain-transfer explanation.","tokens_in":9827,"feed_emoji":"🚗","tokens_out":7928,"duration_ms":71900,"temperature":0.7,"pith_summary":"The paper sets out to show that 3D LiDAR semantic segmentation can be obtained without any 3D annotation and without camera images at inference time. The trick is to align the LiDAR scans of a scene, color each point by its sensor intensity, render many virtual-camera 2D views, and run a 2D segmentation model trained on ordinary street photographs. Each view's labels are projected back onto the 3D points and merged by a voting rule to produce pseudo-labels. On the five static classes considered, the pseudo-labels reach a mean IoU of about 0.49 against ground truth, and training a 3D model on them for unsupervised domain adaptation performs on par with recent methods that use temporal information. If this holds, expensive 3D annotation could be replaced by 2D annotation for static scene classes.","feed_headline":"Intensity-only 2D views can label 3D LiDAR scenes","feed_subtitle":"Zero 3D annotation, no camera at inference: 2D votes over intensity views yield pseudo-labels that train UDA models.","key_machinery":"The load-bearing mechanism is multi-view rendering plus vote fusion. Each aligned LiDAR scene is converted into a large number of greyscale images by projecting points with only their sensor-intensity values into virtual cameras placed near the sensor trajectory. A pretrained 2D segmentation model then produces per-view class logits and hard masks, which are back-projected onto the 3D points and accumulated as votes. Three election estimators are compared: summing hard one-hot votes, summing softmax logits (summing soft votes), and multiplying logits across views (compounding soft votes). Summing soft votes performs best, and the ablation shows that the number of views dominates: dropping from 600 to 100 views cuts mean IoU by more than 86%. The RGB-to-greyscale image augmentation applied when training the 2D model is what makes the camera-trained model usable on intensity renderings.","core_discovery":"The central claim is that direct inference of a 2D segmentation model on intensity-only renderings of dense aligned point clouds is enough to produce useful 3D semantic labels. The paper demonstrates this by coloring each LiDAR point with its measured intensity, clipping and rescaling that intensity, aligning an entire scan sequence into one dense scene, sampling hundreds of virtual camera poses along the sensor trajectory, and running the 2D model on each view. Softmax logits and hard masks are projected back to 3D and accumulated; the sum-of-soft-votes estimator, which totals the logits across views and takes the argmax per point, gives the best results. On the five static classes the best average IoU is about 0.50, with a 0.81 IoU for driveable surface, while building-like manmade structures score worst at 0.37 due to occlusion and limited viewpoint coverage. In the domain-adaptation setting, a 3D model trained on these pseudo-labels reaches 54.2 mean IoU without source pretraining and 61.4 with pretraining, about 6.8 points below a 1% weakly supervised baseline and comparable to the strongest compared UDA method.","pith_inferences":["The paper does not test this, but the weak classes (buildings, vegetation) appear limited by occlusion and viewpoint coverage rather than by 2D recognition quality, so adding depth-aware occlusion masking during back-projection could raise pseudo-label IoU without retraining the 2D model.","Because the RGB-to-greyscale augmentation is already a crude domain transfer, training the 2D model directly on synthetic intensity renderings, or fine-tuning it on a few rendered frames, is a natural extension that could close the remaining domain gap; the paper does not test this.","The vote counts themselves are a per-point confidence signal, so filtering pseudo-labels by agreement across views before UDA training is an untested lever that could improve the 3D model over unfiltered training.","If 2D segmentation models continue to improve, this pipeline should inherit those gains for 3D classes, making the zero-3D-annotation route progressively more attractive without any 3D-specific changes."],"forward_implications":["Static classes such as roads, sidewalks, terrain, and vegetation can be pseudo-labelled across entire LiDAR sequences using only 2D image annotations, removing the need to hand-label 3D point clouds for those classes.","At inference the pipeline needs only LiDAR scans and sensor poses; cameras can be absent, which widens deployment to platforms without camera rigs.","The generated pseudo-labels are strong enough to serve as target-domain supervision for unsupervised domain adaptation, reaching results comparable to a temporal UDA method without using temporal consistency.","Performance scales strongly with the number and diversity of rendered views, so the method's quality is tied to ray-casting effort rather than to 3D network capacity.","The method is explicitly scoped to static classes; moving objects are distorted by scan alignment and are left to future work."],"supporting_citations":[{"why":"Supplies the target LiDAR sequence dataset whose ground-truth labels are used to score pseudo-label quality.","marker":"[2]"},{"why":"Supplies the 2D street-scene image annotations used to train the 2D segmentation model on the five target classes.","marker":"[19]"},{"why":"Provides the transformer-based 2D segmentation architecture that produces the per-view logits used for back-projection.","marker":"[3]"},{"why":"Surveys multi-view score-fusion strategies and supplies the vote estimators compared in the ablation.","marker":"[25]"},{"why":"Provides the temporal UDA method and protocol that serve as the main comparison baseline in the domain-adaptation experiments.","marker":"[6]"},{"why":"Supplies the source-domain 3D labels used to pretrain the 3D model before target pseudo-label fine-tuning.","marker":"[1]"},{"why":"Closest prior approach, using posed camera images rather than generated intensity views; it defines the contrast that motivates the paper's novelty.","marker":"[7]"}],"fun_headline_variants":["2D intensity views label 3D point clouds with no 3D annotation","No 3D labels needed: intensity-only 2D segmentation yields 3D pseudo-labels","From 2D intensity to 3D semantic labels: a camera-free pipeline","Pseudo-labels from 2D intensity maps improve LiDAR UDA without cameras","LiDAR labels via intensity-based 2D views: no 3D annotation required"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on a 2D segmentation model trained on greyscale camera photographs recognizing the same semantic classes in intensity-rendered LiDAR views without any dedicated cross-domain training; if that transfer fails, the method has no labels to offer.","fun_headline_variants_meta":{"raw":{"variants":["2D intensity views label 3D point clouds with no 3D annotation","No 3D labels needed: intensity-only 2D segmentation yields 3D pseudo-labels","From 2D intensity to 3D semantic labels: a camera-free pipeline","Pseudo-labels from 2D intensity maps improve LiDAR UDA without cameras","LiDAR labels via intensity-based 2D views: no 3D annotation required"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001206,"raw_usage":{"total_tokens":5000,"prompt_tokens":1007,"completion_tokens":3993,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":3881}},"tokens_in":623,"tokens_out":3993,"duration_ms":23057,"temperature":1.0,"reasoning_tokens":3881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:54:32.897547+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same aligned scenes and replace each point's real intensity with random values or a flat color, run the same 2D model and voting pipeline, and compare pseudo-label IoU to the reported roughly 0.49; if the score barely drops, the sensor intensity is not what drives the recognition, contradicting the paper's domain-transfer explanation.","supporting_citations":[{"cited_title":"The Mapillary Vistas Dataset for Semantic Understand- ing of Street Scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D street-scene image annotations used to train the 2D segmentation model on the five target classes."},{"cited_title":"Multi-view classification with convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Surveys multi-view score-fusion strategies and supplies the vote estimators compared in the ablation."},{"cited_title":"T-UDA: Temporal Unsupervised Domain Adaptation in Sequential Point Clouds","cited_arxiv_id":null,"evidence_quote":"Provides the temporal UDA method and protocol that serve as the main comparison baseline in the domain-adaptation experiments."},{"cited_title":"Learning 3D Semantic Segmentation with only 2D Image Supervision","cited_arxiv_id":"2110.11325","evidence_quote":"Closest prior approach, using posed camera images rather than generated intensity views; it defines the contrast that motivates the paper's novelty."}],"review_version":1}