{"id":"505b4d59-826c-4351-858e-83b06b2b725b","arxiv_id":"2507.14505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DCHM uses superpixel-based Gaussian Splatting to make monocular depth estimates multiview-consistent, producing point clouds that yield state-of-the-art label-free pedestrian detection on Wildtrack, Terrace, and MultiviewX.","lead":"This paper combines monocular depth estimation with a 3D Gaussian Splatting pipeline so that multiple cameras can build one consistent 3D model of pedestrians, which is then clustered to detect people. It reports the best label-free results on three multiview pedestrian benchmarks, which matters for surveillance and sports analytics because it removes the need for expensive 3D annotations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by fine-tuning on evaluation frames: the paper never states that the 4200 Wildtrack fine-tuning frames exclude the 40 labeled evaluation frames, and the self-training loop makes any overlap consequential.","rationale":"The reader's stated weakest assumption is YOLOv11 mask completeness, but the more decisive issue is the evaluation/fine-tuning overlap. Mask noise would degrade robustness and might lower absolute numbers, but it would not invalidate the comparison if the protocol were clean. A training/evaluation overlap, by contrast, would invalidate the headline numbers themselves, because the model would have been adapted to the test frames during the self-training loop. The paper's Section 4.1 and Section 4.2 describe Wildtrack fine-tuning and evaluation without affirming that the 40 labeled frames are excluded from the 4200 sampled frames, and the 2 fps sampling over the full 35-minute video makes inclusion plausible. The iterative loop in Section 3.1.4 couples pseudo-depth generation, mono-depth fine-tuning, and detection compensation on those same frames, so any overlap directly inflates the reported MODA. This is not an accusation of misconduct; it is a missing protocol detail that must be checked. The reader already gave a CONDITIONAL verdict, and this concern reinforces that condition: the paper should release the exact frame split or the numbers cannot be interpreted as a held-out evaluation. I therefore keep the verdict CONDITIONAL and flag that the split, not mask completeness, is the load-bearing uncertainty.","tokens_in":16118,"tokens_out":6383,"duration_ms":83591,"concrete_test":"Obtain the exact Wildtrack frame indices used for fine-tuning and for the 40-frame evaluation (or rerun with a strict temporal split: reserve a final segment containing all 40 evaluation frames, and use only earlier frames for GS pseudo-depth generation, mono-depth fine-tuning, and detection compensation). If any of the 40 evaluation frames is in the original 4200-frame fine-tuning set, recompute Table 2 on the exclusive split and check whether MODA 84.2 and the 7.6-point gap over UMPD are maintained. Also confirm that MultiviewX and Terrace fine-tuning uses only their declared training images and not validation or evaluation frames.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the state-of-the-art label-free MODA in Table 2 (84.2 on Wildtrack, 80.1 on Terrace, 78.4 on MultiviewX). That claim rests on the evaluation protocol. Section 4.1 says Wildtrack evaluation uses 40 labeled frames; Section 4.2 says 4200 frames per camera sampled at 2 fps from the same 35-minute video are used for mono-depth fine-tuning. The paper never states that the 40 evaluation frames are excluded from the 4200-frame set, and the wording suggests the 4200 are drawn from the whole video, which would include the annotated frames. This matters because the iterative loop in Section 3.1.4 uses those frames for GS pseudo-depth generation, mono-depth fine-tuning, and detection compensation. If the evaluation frames are in that loop, the final monocular depth estimator has been adapted to the exact test images, so the MODA numbers and the comparison with UMPD could reflect memorization rather than a generalizable depth-consistent human model. This is the first thing to verify before interpreting any downstream claim about noise reduction or multiview consistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DCHM, a label-free pipeline for multiview pedestrian detection. The method first generates pseudo-depth labels for pedestrians by optimizing a 3D Gaussian Splatting scene with superpixel-wise supervision, filters those labels by cross-view foreground and depth-consistency checks, and uses the filtered labels to fine-tune a monocular depth estimator. Missed detections are compensated by projecting source-view masks into a reference view and refining them with SAM. At inference, the fine-tuned depth maps are fused into a global Gaussian point cloud, matched across views via a label-matching algorithm, and clustered with DBSCAN to produce pedestrian locations. Experiments on Wildtrack, Terrace, and MultiviewX report label-free state-of-the-art results (MODA 84.2, 80.1, and 78.4 respectively), and the supplementary adds a supervised variant that reaches 95.5 MODA on Wildtrack.","tokens_in":16389,"tokens_out":4384,"duration_ms":55614,"significance":"If the reported numbers are correct under a clean protocol, the paper makes a useful contribution: it demonstrates a fully label-free way to obtain multiview-consistent depth for crowded, sparse-view scenes, and it shows that the resulting point clouds can support not only detection but also multiview pedestrian segmentation. The paper is also commendable for including a decomposition against UMPD, ablations for superpixel-wise optimization, detection compensation, and the iterative loop, plus a DTU-based evaluation of depth quality. The code is promised on a project page. These strengths are, however, conditional on resolving the evaluation-protocol overlap and the self-referential nature of the pseudo-depth loop, as detailed in the major comments.","major_comments":[{"comment":"The Wildtrack evaluation protocol does not state whether the 40 labeled evaluation frames are excluded from the 4200 frames per camera used for mono-depth fine-tuning. Section 4.1 says evaluation uses 40 labeled frames, and Section 4.2 says 4200 images per camera are sampled at 2 fps from the same 35-minute video; the default reading is that the 40 labeled frames lie inside the fine-tuning set. Section 3.1.4 describes an iterative loop in which pseudo-depth generation, mono-depth fine-tuning, and detection compensation all operate on those frames. If the evaluation frames are in the fine-tuning set, the reported MODA values in Tables 2 and 3 and the comparison with UMPD could reflect test-time adaptation rather than a generalizable depth-consistent model. The authors should state whether the 40 frames are excluded, and if they are not, rerun the full pipeline with a clean split.","section":"§4.1, §4.2, §3.1.4"},{"comment":"Several hyperparameters that the pipeline depends on are never reported: the loss weights λ_sp, λ_m, λ_d, λ_o in Eq. (4); the depth-consistency threshold τ in the pseudo-depth filter; the visibility threshold τ_vis in Algorithm 1; the cluster threshold τ_cluster, DBSCAN ϵ and minPts in Algorithm 2; and the blending-weight threshold used in localization. Without these values, the results in Tables 2 and 6 are not reproducible, and the paper provides no sensitivity analysis for parameters that directly control how much pseudo-depth is retained and how many Gaussians form a detection. At minimum, the full parameter setting should be reported, and a sensitivity experiment over τ and τ_vis should be added to show that the state-of-the-art claim is not an artifact of a single threshold choice.","section":"§3.1.1, §3.1.3, §3.2.2, Algorithms 1–4"},{"comment":"The pseudo-depth filtering strategy requires a source-view pixel to reproject into foreground in every reference view and to agree with GS-rendered depth in at least one other view. Supplementary Fig. 10.B reports that 78.67% of Wildtrack pedestrians appear in only one to three views. Pedestrians visible in only one view can never satisfy the depth-consistency condition, and mask errors in any of the few supporting views cause their pixels to be pruned by the foreground filter. Since YOLOv11 masks are the only supervision entering the GS loss and the filtering steps, the paper should quantify mask quality per view and report, for the final Wildtrack detections, how many pedestrians have zero valid pseudo-depth labels, how many detections come from refined multi-view geometry versus compensation, and how sensitive Table 2 is to mask completeness. The limitation statement in §13 acknowledges single-view instability qualitatively, but the main claim needs a quantitative account.","section":"§3.1.1, supplementary Fig. 10.B, §13"},{"comment":"The pseudo-depth loop has a self-referential component that is not discussed. Section 3.1.4 states that the fine-tuned model's updated depth predictions serve as initialization for the next Gaussian optimization cycle, while Section 3.1.1 accepts a pseudo-depth label only if it agrees with GS-rendered depth in another view. Once GS is initialized from the fine-tuned monocular depth, agreement between reprojected depth and GS-rendered depth may simply reflect the initializer rather than independent multiview evidence. The paper should break or analyze this circularity, for example by comparing against fine-tuning with unfiltered monocular depth, by measuring the agreement between pseudo-labels and an independent source (e.g., a stereo method or ground-truth depth on a subset), or by reporting how much of the improvement remains after only one loop without re-initialization.","section":"§3.1.4, §3.1.1, Fig. 5"}],"minor_comments":[{"comment":"The word 'Supersingly' should be 'Surprisingly'.","section":"§4.3"},{"comment":"'scaler weights' should be 'scalar weights' in the description of Eq. (4).","section":"§3.1.1"},{"comment":"The Terrace dataset is cited as reference [12], but reference [12] is the probabilistic occupancy map paper; the correct Terrace dataset citation should be provided.","section":"References"},{"comment":"Several baselines have '-' entries for Terrace and MultiviewX; the paper should state whether those results are unavailable from the original papers or were omitted for another reason, since the comparison is uneven otherwise.","section":"Table 2"},{"comment":"Report whether the numbers in Table 2 are single runs or averaged over multiple runs, and if averaged, give standard deviations; if the pipeline is deterministic, say so explicitly.","section":"§4.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The most consequential issue is the potential overlap between the Wildtrack fine-tuning frames and the 40 evaluation frames. This should be resolved before the paper is considered further, because it directly affects the headline numbers. The circularity of the pseudo-depth loop and the unquantified dependence on YOLOv11 mask quality are also substantive and should be addressed with additional analysis. The paper otherwise fits the journal and the proposed method is interesting if the protocol is clean."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering contribution to label-free multiview pedestrian detection, with large reported gains over UMPD. The main thing I'd check before trusting the headline numbers is the Wildtrack protocol: the paper says 4200 frames per camera are sampled for fine-tuning and 40 labeled frames are used for evaluation, but never states that the 40 are excluded from the 4200. If they overlap, the fine-tuned depth model has seen the test images, and the 84.2 MODA could be inflated by memorization. This needs a one-sentence clarification, not a redesign.\n\nWhat's genuinely new: the superpixel-wise Gaussian splatting initialization makes sparse-view GS work where SfM fails, and the cross-view pseudo-depth filtering is a reasonable way to keep only geometrically consistent labels. The iterative loop that alternates GS-based pseudo-labeling, mono-depth fine-tuning, and SAM-based detection compensation is clever, and the paper shows with ablations that each piece helps. The decomposition against UMPD is fair, and the supplementary's DTU depth evaluation is a useful external check that the pseudo-labels carry real geometric signal, not just self-consistency.\n\nSoft spots: besides the split ambiguity, there are no error bars or significance tests—MODA differences of 7-10 points are probably real, but we don't know the run-to-run variance. Many thresholds and loss weights are unstated, and no code is linked in the paper itself (the project page is mentioned, but I couldn't verify the code is there). The pseudo-depth loop is self-referential; the DTU experiment mitigates that concern, but the authors should show a few qualitative or quantitative checks that the refined depth is closer to true metric depth, not just closer to the GS renders.\n\nWho this is for: anyone working on multiview pedestrian detection or deploying 3DGS in sparse outdoor scenes. It deserves a serious referee: the idea is clearly articulated, the experiments are broad, and the issues are fixable with reporting changes. I'd send it to review with a request to clarify the Wildtrack split, add variance estimates, and release hyperparameters.\n\nIf I work in this space, I'd cite it once the split is confirmed, but I wouldn't use the numbers as a baseline until then.","headline":"Plausible and useful label-free multiview detection pipeline, but the Wildtrack train/test split ambiguity undercuts the headline numbers until resolved.","tokens_in":16951,"tokens_out":3184,"would_cite":false,"duration_ms":36052,"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":"DCHM claims that a label-free pipeline built on superpixel-wise Gaussian splatting and refined monocular depth can detect pedestrians in sparse-view crowded scenes with MODA 84.2 on Wildtrack, 80.1 on Terrace, and 78.4 on MultiviewX.","keywords":["multiview pedestrian detection","label-free detection","depth consistency","Gaussian splatting","pseudo-depth learning","superpixel initialization","multiview segmentation","pedestrian localization"],"falsifier":"Run the pipeline on a scene with ground-truth depth and masks, then measure the precision of the retained pseudo-depth labels: if many labels surviving cross-view filtering exceed a small depth-error tolerance, or if randomly dropping 20% of per-view masks leaves MODA unchanged, the claimed dependence of the geometry on complete masks would be falsified.","tokens_in":15884,"feed_emoji":"🚶","tokens_out":7522,"duration_ms":78036,"temperature":0.7,"pith_summary":"Multiview pedestrian detection needs to fuse sparse camera views into one consistent 3D picture, and this paper argues that the bottleneck is depth consistency, not annotations. DCHM is a fully label-free pipeline: it initializes Gaussian splatting with per-pedestrian superpixels, optimizes it against segmentation masks to obtain pseudo-depth, filters those depths across views, fine-tunes a monocular depth network on them, and repeats the loop. The authors claim this yields point clouds precise enough that simple clustering recovers pedestrians, reaching MODA 84.2 on Wildtrack, 80.1 on Terrace, and 78.4 on MultiviewX without any 3D labels. They also claim it is the first method to reconstruct pedestrians and perform multiview segmentation in sparse-view, crowded scenes. If correct, the contribution matters because it removes the costly 3D annotation requirement and makes detection geometry accurate enough to later feed supervised localization, reaching 95.5 MODA on Wildtrack.","feed_headline":"Label-free pipeline hits 84.2 MODA on Wildtrack","feed_subtitle":"Superpixel Gaussian splatting fine-tunes monocular depth so sparse camera views fuse into one point cloud.","key_machinery":"The load-bearing machinery is superpixel-wise Gaussian Splatting used as a pseudo-depth generator. Gaussian splatting normally needs good 3D initialization and densely overlapping views, so the paper replaces Structure-from-Motion with a uniform sampling plus filtering initialization: rays through superpixel centers are sampled uniformly, with Gaussian scale fixed by superpixel area and ray distance. A combined loss enforces superpixel-wise photometric agreement, mask agreement, opacity near zero or one, and small depth variance inside each pedestrian mask. Two geometric filters then decide which rendered depths become labels: cross-view foreground filtering requires reprojection into the foreground in every reference view, and cross-view depth consistency filtering requires agreement with at least one other view's rendered depth. A second machinery piece is multiview label matching, which propagates Gaussian IDs through masks view by view using blending weights, so the same pedestrian gets one ID across cameras and its Gaussians can be clustered.","core_discovery":"The central discovery is that cross-view depth consistency can be self-learned from segmentation masks alone, without 3D annotations, and that this consistency is what makes multiview human modeling usable. Concretely, for each pedestrian the method builds K superpixels, initializes a cloud of Gaussians by ray-sampling through superpixel centers, and optimizes with a loss combining superpixel photometric error, mask overlap, opacity regularization, and within-mask depth flatness. Because Gaussian splatting needs multiview support, only pixels that reproject into the foreground in every view and whose depth agrees with another view's rendered depth are kept as pseudo-depth labels; those labels fine-tune a monocular depth network, and the network's improved depth reinitializes the next splatting round. At inference, per-view masks are matched across views by assigning shared Gaussian IDs, and DBSCAN clustering of each ID's Gaussians gives detections. The claimed result is state-of-the-art label-free detection with higher localization precision, plus the first 3D pedestrian reconstruction and multiview segmentation under sparse, crowded, wide-baseline conditions.","pith_inferences":["An editorial inference: nothing in the depth-consistency loop is pedestrian-specific, so the same superpixel Gaussian pseudo-depth fine-tuning could be applied to other sparse-view multiview detection tasks, such as vehicles or animals, whenever per-view masks are available.","The paper's own limitation note implies that detection-compensation quality is capped by depth accuracy; a natural extension would be to gate SAM prompts with an estimated per-projection depth uncertainty and to validate the compensated masks before accepting them.","Because 78.67% of Wildtrack pedestrians appear in only one to three views, the method's gains likely concentrate on the multi-view majority of pixels; incorporating temporal consistency is the most direct test of whether the remaining single-view-only cases can be pushed further."],"forward_implications":["DCHM reports the best label-free results on the three benchmarks: MODA 84.2 on Wildtrack, 80.1 on Terrace, and 78.4 on MultiviewX, with Wildtrack MODP of 80.3 versus 61.2 for the previous label-free baseline UMPD.","When the same human modeling is wrapped with the supervised localization head used in reference [31], it reaches 95.5 MODA on Wildtrack at 6.1 FPS, so the learned geometry transfers to annotation-based pipelines as well.","The iterative loop converges in three rounds: after the third optimization cycle, valid pseudo-depth regions stop growing and accuracy gains become minimal, giving a practical stopping rule.","Because no 3D annotations are needed, the same training recipe can be applied to a new calibrated camera setup using only images and per-view segmentation masks, as done for the 4200 sampled Wildtrack frames per camera.","At inference, human modeling takes about 0.1 seconds per frame and full label-free detection runs at 1.2 FPS, a real-time operating point for surveillance-scale use."],"supporting_citations":[{"why":"Supplies the Gaussian splatting representation and renderer that the pipeline optimizes, prunes, and grows to produce pseudo-depth.","marker":"[18]"},{"why":"Provides the monocular depth network that is fine-tuned on pseudo-depth labels and reused at inference for consistent depth maps.","marker":"[50]"},{"why":"Produces the per-view pedestrian segmentation masks used in the Gaussian splatting loss, foreground filtering, and label matching.","marker":"[19]"},{"why":"UMPD is the previous label-free baseline whose human modeling is compared and replaced; the method is evaluated against its MODA and MODP numbers.","marker":"[27]"},{"why":"SAM turns projected box-point prompts from other views into compensatory masks for pedestrians missed by the single-view detector.","marker":"[20]"},{"why":"DBSCAN clusters the ID-assigned Gaussians into final pedestrian locations, with the number of Gaussians per cluster serving as confidence.","marker":"[10]"},{"why":"SLIC generates the superpixels that define superpixel-wise supervision and set the color and scale of the initialized Gaussians.","marker":"[1]"},{"why":"Structure-from-Motion is identified as failing under wide camera baselines, motivating the ray-uniform-sampling initialization that the method uses instead.","marker":"[35]"}],"fun_headline_variants":["Depth-consistent Gaussian splatting lifts multiview detection","Self-learned depth from masks boosts multiview pedestrian detection","No 3D labels needed: DCHM fuses views via depth consistency","Superpixel splatting achieves cross-view depth for sharper detection","First 3D pedestrian reconstruction from sparse multiview without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-view pedestrian masks produced by YOLOv11 are complete and accurate enough to supervise every downstream stage; mask misses or over-segmentation in the many pedestrians seen by only one to three views would directly corrupt which Gaussians survive, which IDs get assigned, and whether neighboring people merge.","fun_headline_variants_meta":{"raw":{"variants":["Depth-consistent Gaussian splatting lifts multiview detection","Self-learned depth from masks boosts multiview pedestrian detection","No 3D labels needed: DCHM fuses views via depth consistency","Superpixel splatting achieves cross-view depth for sharper detection","First 3D pedestrian reconstruction from sparse multiview without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1802,"prompt_tokens":978,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":735}},"tokens_in":594,"tokens_out":824,"duration_ms":523581,"temperature":1.0,"reasoning_tokens":735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:54:36.532568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a scene with ground-truth depth and masks, then measure the precision of the retained pseudo-depth labels: if many labels surviving cross-view filtering exceed a small depth-error tolerance, or if randomly dropping 20% of per-view masks leaves MODA unchanged, the claimed dependence of the geometry on complete masks would be falsified.","supporting_citations":[{"cited_title":"Un- supervised multi-view pedestrian detection","cited_arxiv_id":null,"evidence_quote":"UMPD is the previous label-free baseline whose human modeling is compared and replaced; the method is evaluated against its MODA and MODP numbers."},{"cited_title":"Segment anything","cited_arxiv_id":null,"evidence_quote":"SAM turns projected box-point prompts from other views into compensatory masks for pedestrians missed by the single-view detector."},{"cited_title":"Density-based spatial clustering of applications with noise","cited_arxiv_id":null,"evidence_quote":"DBSCAN clusters the ID-assigned Gaussians into final pedestrian locations, with the number of Gaussians per cluster serving as confidence."},{"cited_title":"Slic superpixels","cited_arxiv_id":null,"evidence_quote":"SLIC generates the superpixels that define superpixel-wise supervision and set the color and scale of the initialized Gaussians."},{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Structure-from-Motion is identified as failing under wide camera baselines, motivating the ray-uniform-sampling initialization that the method uses instead."}],"review_version":1}