{"id":"ba8eb3f4-5652-4323-9be3-052c1085226b","arxiv_id":"2412.04117","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Mean-teacher self-training with local-max pseudo-labeling adapts multi-view pedestrian detectors across camera rigs without external labeled monocular data, improving MODA from 70.0 to 85.4 on MultiviewX to Wildtrack.","lead":"This paper adapts multi-view pedestrian detectors to new camera rigs without new labels, using a mean-teacher self-training loop with a local-maximum-based pseudo-label filter. It reports large gains on benchmarks such as Wildtrack and MultiviewX, with an important caveat about how thresholds were chosen.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline MODA gain is carried by test-set threshold selection: tau is swept on target test data (Sec. 4.2) and pseudo-label thresholds are chosen per benchmark from final MODA (Sec. 4.5); fixed-threshold re-evaluation is required before the SOTA claim can stand.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing issue: thresholds are selected using target test labels, so the reported numbers are optimistic and the comparison with prior fixed-threshold methods is unfair. My read confirms this; Table 4 alone shows a swing of roughly 30 MODA on Wildtrack→MultiviewX depending on tau, which is larger than the margin over several competing methods. The ImageNet-pretraining contradiction is real and should be corrected in the text, but it does not directly alter the empirical magnitudes the way the threshold protocol can. Therefore, if a fixed-threshold rerun preserves most of the gain, the paper is a solid empirical contribution; if it does not, the SOTA and 'large margin' claims should be withdrawn. The reader's CONDITIONAL verdict remains the right call, so no verdict change is needed.","tokens_in":17696,"tokens_out":8087,"duration_ms":83958,"concrete_test":"Rerun the two cross-domain benchmarks with one fixed tau=0.4 for both pseudo-label generation and final evaluation (the protocol used by MVDet and GMVD), and also with the pseudo-label threshold selected on a source-domain validation split or on unlabeled target statistics rather than from target test MODA. Report MVUDA and Lima et al. under identical threshold rules. If Wildtrack→MultiviewX falls from 82.4 to near the 35.9 baseline or below the comparison method, the headline SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the evaluation and threshold protocol, not the self-training idea. The quantitative anchor of the paper is the Wildtrack→MultiviewX jump from 35.9 to 82.4 MODA. Section 4.2 states that the final evaluation threshold tau is swept over {0.05,...,0.95} and the value with highest MODA is reported, and Section 4.5 sets each benchmark's pseudo-label threshold from Table 4, which is generated from target test predictions. Both choices use target ground truth. Table 4 shows how much this matters: with local-max pseudo-labeling and no augmentation, Wildtrack→MultiviewX achieves 73.2 MODA at tau=0.2, 68.7 at tau=0.3, and 43.5 at tau=0.4, while MultiviewX→Wildtrack achieves 75.8 at tau=0.4 but 70.8 at tau=0.3. Prior methods are evaluated at fixed tau=0.4, so the headline comparison with Lima et al. is not on equal terms. The method may still be useful, but the magnitude of the claimed improvement is entangled with an oracle threshold choice and needs to be re-derived under a fixed, pre-specified protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MVUDA, an unsupervised domain adaptation method for multi-view pedestrian detection. It uses a mean-teacher self-training framework in which a student detector is trained on labeled source data and pseudo-labeled target data, with pseudo-labels generated by a teacher and post-processed by a newly proposed 'local-max' procedure that retains only local maxima of the predicted occupancy map above a confidence threshold. Experiments on Wildtrack, MultiviewX, and GMVD report large gains over the source-only baseline, close-to-oracle performance on several benchmarks, and state-of-the-art results among methods that do not use auxiliary labeled data. The paper also introduces two new cross-scene benchmark splits, GMVD1→MultiviewX and GMVD2→MultiviewX.","tokens_in":17953,"tokens_out":7064,"duration_ms":67913,"significance":"If the reported numbers hold under a valid evaluation protocol, this is a useful contribution: it removes the dependence on external monocular detectors that characterizes prior UDA work by Lima et al., and it proposes a simple, well-motivated pseudo-label post-processing idea that appears to improve self-training. The paper is unusually thorough on ablations: Tables 3, 4, 6-8 and supplementary Tables 9-13 cover the main components (mean teacher, augmentation, pseudo-labeling variants, α, λ, kd, training length). The qualitative claim that self-training plus local-max pseudo-labeling helps is supported by these ablations. However, the quantitative headline claims are currently entangled with test-set threshold selection: both the evaluation threshold and the pseudo-label threshold are chosen using target-domain ground truth. Therefore the magnitude of the claimed improvement over prior methods is not yet established.","major_comments":[{"comment":"The evaluation protocol selects the detection threshold on the target test set: 'we evaluate the model on the range τ ∈ {0.05, 0.10, ..., 0.95} and select the result with highest MODA.' This is an oracle selection on test labels. Because prior methods are evaluated at fixed τ=0.4 (as the paper itself notes), the comparisons in Tables 1 and 2 are not on equal terms. The Wildtrack→MultiviewX headline jump from 35.9 to 82.4 is reported at τ=0.2; Table 4 shows that without augmentation the same method at τ=0.4 gives 43.5 and at τ=0.3 gives 68.7. Please re-evaluate all methods under a fixed, pre-specified threshold (at least τ=0.4) and, additionally or alternatively, choose τ without using target labels (e.g., on a labeled source validation set).","section":"Sec. 4.2"},{"comment":"The pseudo-label threshold is also chosen per benchmark from the final MODA table: the text states the threshold is set 'following the experiments presented in Tab. 4,' and Table 4 is computed from target test predictions. Since the pseudo-label threshold directly controls the training targets for self-training, the entire pipeline is tuned to the target test ground truth. This is a second use of test labels for model selection. Please either fix the pseudo-label thresholds a priori (e.g., using source-domain validation statistics) or report a sensitivity analysis showing that the conclusions are stable across a range of reasonable thresholds.","section":"Sec. 4.5 and Table 4"},{"comment":"The claim that local-max pseudo-labeling is 'robust' to the threshold τ is not fully supported by the data. On Wildtrack→MultiviewX, local-max pseudo-labeling yields 73.2 MODA at τ=0.2, 68.7 at τ=0.3, and 43.5 at τ=0.4. A drop of ~30 MODA over 0.2 units of threshold is a strong sensitivity, not robustness. The text also says the method 'yields improvements over the baseline for a wider range of τ,' but on this benchmark the range of τ for which the improvement is large is narrow. Please qualify this claim with the actual range and report results at fixed thresholds.","section":"Sec. 4.5, Table 4"}],"minor_comments":[{"comment":"The benchmark name is inconsistent: Table 2 lists 'MultiviewX 1,2,6→3,4,5' while Table 4 lists 'MultiviewX 1,2,3→4,5,6'. Please align the camera indices.","section":"Table 4 vs Table 2"},{"comment":"The text says λ=1.0 yields 'the best overall performance (tied with λ=2.0),' but in Table 10 λ=1.0 gives 79.7 vs λ=2.0 giving 79.1 on MultiviewX→Wildtrack, and 87.8 vs 88.4 on GMVD1→MultiviewX. Please clarify which metric supports the 'best overall' statement.","section":"Supplementary, Sec. 5.3"},{"comment":"The paper states that the purpose of selecting the best MODA over τ is to 'ensure that the experimental results are not affected by the specific choice of τ.' In fact, this procedure makes the results depend on the test set through the chosen threshold; a sentence acknowledging this distinction would be more accurate.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the threshold-selection protocol. I do not see an internal inconsistency in the method itself; the local-max idea is plausible and the ablations are extensive. The revision should focus on re-running the evaluation with a fixed, pre-specified protocol and reporting the sensitivity of the headline numbers. If the fixed-threshold numbers remain strong, the paper would be a solid empirical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible UDA paper for multi-view pedestrian detection, and the local-max pseudo-labeling idea is worth knowing. The paper removes the external monocular detector that Lima et al. relied on, adds a local-max filter to pseudo-label post-processing, and introduces two new GMVD-to-MultiviewX benchmarks. The local-max idea is simple, and the ablations (Tables 4-8 and the supplementary) support that it yields more reliable detections at low confidence thresholds. The writing is clear and the experiments are thorough.\n\nThe soft spot is the evaluation protocol, and it is load-bearing. Section 4.2 sweeps the evaluation tau over the test set and reports the best MODA. Section 4.5 then sets each benchmark's pseudo-label tau from the final MODA table. Prior methods like Lima et al. use a fixed tau=0.4, so the comparison is not on equal terms. Table 4 shows the sensitivity: Wildtrack->MultiviewX goes from 73.2 at tau=0.2 to 68.7 at tau=0.3, and MultiviewX->Wildtrack from 75.8 at tau=0.4 to 70.8 at tau=0.3. That said, at any reasonable threshold the local-max variant beats vanilla and the baseline by a wide margin on Wildtrack->MultiviewX, so the method is not an artifact of the selection. The authors should re-report under a fixed, pre-specified protocol (e.g., tau=0.4 everywhere, or a validation split) before claiming SOTA.\n\nMinor points: the 'eliminates the need for auxiliary labeled datasets' claim overstates things because the backbone is ImageNet-pretrained, which is still external labeled data. Also, no code or multiple seeds are provided, which adds uncertainty to the exact numbers, and there is a typo in Table 4's heading ('MultiviewX 1,2,3'). These are fixable.\n\nWho it's for: researchers working on multi-view detection or UDA for detection; the threshold protocol discussion is a useful cautionary example. It deserves a serious referee. I would send it out, with the request that the authors fix the protocol and re-evaluate all comparisons at common thresholds.","headline":"Solid UDA method with a genuinely useful pseudo-labeling trick, but the headline numbers depend on test-set threshold selection and need re-evaluation under a fixed protocol.","tokens_in":18541,"tokens_out":3565,"would_cite":false,"duration_ms":35082,"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":"MVUDA shows that a multi-view pedestrian detector can be adapted to a new camera rig using only unlabeled target images, eliminating the need for auxiliary labeled datasets.","keywords":["unsupervised domain adaptation","multi-view pedestrian detection","mean teacher self-training","pseudo-labeling","bird's-eye view occupancy map","cross-camera rig adaptation","Wildtrack","MultiviewX"],"falsifier":"Run the two cross-domain benchmarks with $\\tau$ fixed to 0.4 for both pseudo-label generation and final evaluation, without any test-set sweep, and check whether MVUDA still reaches the reported 82.4 MODA on Wildtrack→MultiviewX and 85.4 on MultiviewX→Wildtrack; the paper's own ablation shows local-max self-training ranging from 73.2 at $\\tau=0.2$ to 43.5 at $\\tau=0.4$ on the former, so a fixed threshold would materially change the headline result.","tokens_in":17461,"feed_emoji":"🚶","tokens_out":12124,"duration_ms":97847,"temperature":0.7,"pith_summary":"MVUDA tries to establish that a multi-view pedestrian detector trained on labeled images from one camera rig can be adapted to a different rig using only unlabeled images from the target rig, with no auxiliary labeled monocular data. The paper claims this is the first strict unsupervised domain adaptation (UDA) method for multi-view pedestrian detection that eliminates the need for external labeled datasets. The method combines mean teacher self-training with a pseudo-labeling step that only accepts detections at local maxima of the predicted bird's-eye-view occupancy map. On the Wildtrack→MultiviewX benchmark, MVUDA reports raising MODA from 35.9 to 82.4, outperforming the prior UDA method [28] even though that method relied on a monocular detector trained on large labeled datasets. If the claims hold, multi-view pedestrian detection becomes practical to deploy on new camera setups without collecting target-domain annotations.","feed_headline":"No target labels: multi-view pedestrian detection adapts to new rigs","feed_subtitle":"MVUDA lifts cross-camera detection from 35.9 to 82.4 MODA with no target-domain labels.","key_machinery":"The method builds on the multi-view detector of [43], a variant of [20]: a ResNet-18 (a standard convolutional image-feature extractor) produces per-view features, a perspective transformation projects these into a shared bird's-eye-view (BEV) grid, average pooling fuses the views, and dilated convolutions regress a probabilistic occupancy map (a per-cell score of pedestrian presence on the ground plane). The mean teacher [40] is an exponential moving average of the student weights; it generates pseudo-labels on unaugmented target images while the student trains on strongly augmented images, using a Gaussian-softened MSE loss on both source labels and target pseudo-labels. The novel component is local-max pseudo-labeling: a candidate detection is accepted only if its score is at least as large as every score in a surrounding square neighborhood defined by $k_d$ (a 70 cm by 70 cm square when $k_d=3$) and also exceeds a threshold $\\tau$, so non-maximum points are never turned into pseudo-labels. Dropview [43] and 3D random occlusion [32] augmentations are used in both the baseline and self-training.","core_discovery":"The central claim is that strict unsupervised domain adaptation is achievable for multi-view pedestrian detection without any external labeled datasets or pretrained monocular detectors. The paper argues that mean teacher self-training works well when pseudo-labels are restricted to local maxima of the teacher's predicted occupancy map: vanilla threshold-and-NMS post-processing can select detections that are not local maxima, and these are unreliable exactly when the confidence threshold is ambiguous in a new domain. The proposed local-max pseudo-labeling yields reliable pseudo-labels at lower confidence levels, which matters most when the source-only baseline transfers poorly, as in Wildtrack→MultiviewX where the baseline scores 35.9 MODA and MVUDA reports 82.4. The paper also introduces two cross-scene camera-rig benchmarks (GMVD1→MultiviewX and GMVD2→MultiviewX) and reports the highest MODA among methods that do not use auxiliary labeled data on all studied benchmarks.","pith_inferences":["Editorial inference: The reported numbers should be read as upper bounds because the paper selects the detection threshold by sweeping $\\tau$ over {0.05, ..., 0.95} on the test set and keeping the best MODA, and it sets the self-training pseudo-label threshold per benchmark from the final results table.","Editorial inference: A conservative rerun with $\\tau$ fixed a priori (or chosen on a held-out validation split) would likely shrink the headline gains on some benchmarks, since the paper's no-augmentation ablation shows Wildtrack→MultiviewX local-max self-training varying from 73.2 MODA at $\\tau=0.2$ to 43.5 at $\\tau=0.4$.","Editorial inference: Local-max pseudo-labeling is a generic idea that could transfer to other dense-prediction UDA problems where the model outputs a confidence map expected to peak at object centers, such as keypoint detection or semantic segmentation.","Editorial inference: Because the method removes the dependence on external labeled datasets, a natural testable extension is applying the same pipeline to other object classes (e.g., vehicles) using a source rig and a different target rig, where no monocular pretrained detector exists."],"forward_implications":["A multi-view pedestrian detector can be adapted to a new camera rig with only unlabeled target video, removing the need for expensive multi-view annotation or for an external monocular detector.","The largest gains occur when the source-only model is weak: on Wildtrack→MultiviewX, MVUDA raises MODA from 35.9 to 82.4, so self-training can largely rescue a detector that transfers poorly across rigs.","Local-max pseudo-labeling makes self-training more robust to the choice of confidence threshold, outperforming vanilla pseudo-labeling on most benchmarks and improving over baseline over a wider range of $\\tau$.","On the new cross-scene benchmarks GMVD1→MultiviewX (89.0 MODA) and GMVD2→MultiviewX (88.8), MVUDA approaches the Oracle trained with target labels (91.2), indicating that most of the benefit of target labels is recoverable without them.","The framework establishes a UDA baseline that does not depend on any auxiliary labeled dataset, which the paper argues is important for practical applications and for extending to new object types."],"supporting_citations":[{"why":"Supplies the multi-view detector architecture and the Dropview augmentation that MVUDA builds on, and defines two Wildtrack camera-rig split benchmarks.","marker":"[43]"},{"why":"Supplies the perspective-transformation BEV detector, the Gaussian-smoothed MSE training targets, and the evaluation convention for multi-view pedestrian detection.","marker":"[20]"},{"why":"Is the prior UDA method for multi-view pedestrian detection that relies on an external monocular detector, and is the main baseline MVUDA outperforms on Wildtrack→MultiviewX.","marker":"[28]"},{"why":"Provides the mean teacher weight-averaging mechanism that generates pseudo-labels during self-training.","marker":"[40]"},{"why":"Provides the 3D random occlusion augmentation used in the self-training framework.","marker":"[32]"},{"why":"Provides the Wildtrack real-world dataset used as source or target in the cross-domain benchmarks.","marker":"[7]"},{"why":"Supplies the external monocular 3D pedestrian detector that [28] depends on, illustrating the auxiliary labeled-data reliance MVUDA removes.","marker":"[26]"},{"why":"Provides the MV Aug two-level augmentation that the paper compares and ultimately excludes from the final configuration.","marker":"[11]"}],"fun_headline_variants":["From 35.9 to 82.4 MODA: multi-view detection adapts to new rigs","No labels, no external data: MVUDA tops cross-rig benchmarks","Local-max pseudo-labels power UDA for multi-view pedestrian detection","Unsupervised adaptation: multi-view detection hits 82.4 MODA on new rig","MVUDA: cross-camera detection without target labels or monocular data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers depend on selecting the detection and pseudo-label thresholds from the target-domain test labels (the paper sweeps $\\tau$ and keeps the value with the highest MODA detection-accuracy score, then sets per-benchmark $\\tau$ from the final results table), so the reported gains assume the practitioner can peek at the test set before fixing thresholds.","fun_headline_variants_meta":{"raw":{"variants":["From 35.9 to 82.4 MODA: multi-view detection adapts to new rigs","No labels, no external data: MVUDA tops cross-rig benchmarks","Local-max pseudo-labels power UDA for multi-view pedestrian detection","Unsupervised adaptation: multi-view detection hits 82.4 MODA on new rig","MVUDA: cross-camera detection without target labels or monocular data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001446,"raw_usage":{"total_tokens":5810,"prompt_tokens":918,"completion_tokens":4892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":4783}},"tokens_in":534,"tokens_out":4892,"duration_ms":39590,"temperature":1.0,"reasoning_tokens":4783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:44:04.653527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two cross-domain benchmarks with $\\tau$ fixed to 0.4 for both pseudo-label generation and final evaluation, without any test-set sweep, and check whether MVUDA still reaches the reported 82.4 MODA on Wildtrack→MultiviewX and 85.4 on MultiviewX→Wildtrack; the paper's own ablation shows local-max self-training ranging from 73.2 at $\\tau=0.2$ to 43.5 at $\\tau=0.4$ on the former, so a fixed threshold would materially change the headline result.","supporting_citations":[{"cited_title":"Bringing generalization to deep multi-view pedestrian detection","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-view detector architecture and the Dropview augmentation that MVUDA builds on, and defines two Wildtrack camera-rig split benchmarks."},{"cited_title":"Multi- view detection with feature perspective transformation","cited_arxiv_id":null,"evidence_quote":"Supplies the perspective-transformation BEV detector, the Gaussian-smoothed MSE training targets, and the evaluation convention for multi-view pedestrian detection."},{"cited_title":"Mean teacher for unsupervised domain adaptation in multi-view 3D pedestrian detection","cited_arxiv_id":null,"evidence_quote":"Is the prior UDA method for multi-view pedestrian detection that relies on an external monocular detector, and is the main baseline MVUDA outperforms on Wildtrack→MultiviewX."},{"cited_title":"3D random occlusion and multi-layer projection for deep multi-camera pedestrian localization","cited_arxiv_id":null,"evidence_quote":"Provides the 3D random occlusion augmentation used in the self-training framework."},{"cited_title":"Wildtrack: A multi-camera HD dataset for dense unscripted pedes- trian detection","cited_arxiv_id":null,"evidence_quote":"Provides the Wildtrack real-world dataset used as source or target in the cross-domain benchmarks."},{"cited_title":"3D pedestrian localization using mul- tiple cameras: A generalizable approach","cited_arxiv_id":null,"evidence_quote":"Supplies the external monocular 3D pedestrian detector that [28] depends on, illustrating the auxiliary labeled-data reliance MVUDA removes."},{"cited_title":"Two-level data augmentation for calibrated multi-view de- tection","cited_arxiv_id":null,"evidence_quote":"Provides the MV Aug two-level augmentation that the paper compares and ultimately excludes from the final configuration."}],"review_version":1}