{"id":"ae222139-79cb-4969-87ea-4c1b6bdcf421","arxiv_id":"2412.14571","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SCKD uses a LiDAR-radar fusion teacher and semi-supervised output distillation to train a radar-only student that outperforms prior radar-only methods on the VoD and ZJUODset benchmarks.","lead":"This paper trains a radar-only 3D object detector by distilling knowledge from a LiDAR-radar fusion teacher, using semi-supervised pseudo-labels instead of ground truth. The method improves radar-only detection accuracy on the View-of-Delft dataset while keeping real-time inference, and it can exploit extra unlabeled data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SSOD threshold σ=0.1 is never ablated; if the final mAP is sensitive to this threshold, the reported 10.38% gain and the 'SSOD > GT' claim lack robustness.","rationale":"The reader identified pseudo-label quality as the weakest assumption. I agree with the general concern but sharpen it to a specific, testable issue: the confidence threshold σ is selected at 0.1 without any ablation. The paper's own ablation (Table 3) supports SSOD at this single threshold, but it does not establish that the result is stable. If the threshold is sensitive, the 10.38% gain could be a tuned artifact. This concern does not overturn the reader's verdict—the paper deserves conditional acceptance pending clarification—but it points to a concrete experiment that would settle it. I set verdict_should_be to UNCHANGED because the conditional verdict already captures the need for additional verification. The agreement is partial because the reader's formulation ('teacher outputs provide supervision of sufficient quality') is broader than the specific threshold-sensitivity issue I emphasize.","tokens_in":12968,"tokens_out":15220,"duration_ms":125647,"concrete_test":"Re-run the full SCKD pipeline on the VoD dataset with σ ∈ {0.05, 0.2, 0.3} while keeping all other hyperparameters fixed, and report validation mAP (entire annotated area and driving corridor). Additionally, compute the teacher's precision and recall against ground truth on the validation set for each σ. If the student's mAP varies by more than 1 point across this range, the choice of σ is load-bearing and the robustness claim fails; if the mAP is flat, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the Semi-Supervised Output Distillation (SSOD) component, which replaces ground-truth supervision with teacher detections selected by Eq. (13) at a confidence threshold σ=0.1. This threshold is very low, so many low-confidence and likely false-positive proposals become training targets for the student. The paper's ablation (Table 3, rows (b) vs (a)) shows SSOD alone outperforms GT-only training, but only at this single threshold. No sensitivity analysis of σ is presented, and the precision/recall of the teacher's pseudo-labels is not reported. If the final mAP depends strongly on σ, the reported 10.38% improvement over the SECOND baseline—and the broader claim that SSOD is 'more effective than GT'—may be an artifact of tuning σ rather than a robust property of the framework. This is load-bearing because SSOD is the sole output supervision for the student in the main VoD experiments; if the pseudo-labels are too noisy, the student inherits the teacher's systematic errors, and the advantage over state-of-the-art radar-only methods may not generalize to other datasets or operating conditions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SCKD, a semi-supervised cross-modality knowledge distillation method for 4D radar 3D object detection. A Lidar-radar fusion teacher network, enhanced by an adaptive fusion module with random modality dropout, is used to train a radar-only SECOND student. Knowledge is transferred through two feature-distillation losses (LRFD and FRFD) and a semi-supervised output distillation loss (SSOD) in which teacher detections above a confidence threshold sigma replace ground-truth labels as student supervision. Experiments on VoD and ZJUODset report that the radar-only student improves mAP by 10.38 points over the SECOND baseline on VoD, surpasses previous radar-only methods, and approaches or exceeds most radar-camera fusion methods while running at 39.3 FPS.","tokens_in":13223,"tokens_out":5616,"duration_ms":46305,"significance":"If the empirical claims are robust, SCKD is a useful contribution: it demonstrates that a radar-only detector can benefit from a Lidar-radar teacher at no inference cost, and it provides a semi-supervised mechanism for exploiting unlabeled radar data. The public code release, evaluation on two datasets, and the teacher-design analysis (Table 5) are strengths. The main limitations are that all numbers come from single runs, the reported margin over the strongest radar-only competitor (SMURF, Table 1) is only 1.11 mAP, and the key pseudo-label threshold sigma is not ablated. These issues do not invalidate the idea, but they need to be addressed before the state-of-the-art and 'SSOD > GT' claims can be considered established.","major_comments":[{"comment":"The central comparison in the ablation, row (b) vs row (a), is used to claim that SSOD alone is more effective than ground-truth supervision, yet this comparison is only performed at sigma=0.1. The threshold is not ablated, and the precision/recall of the teacher's pseudo-labels is not reported. Because sigma defines the entire supervision signal for the student in the main VoD configuration, a sensitivity analysis over sigma (and ideally pseudo-label quality statistics) is required to show that the reported 10.38-point gain is not an artifact of this particular threshold.","section":"SSOD: Semi-Supervised Output Distillation, Eq. (13), Table 3"},{"comment":"Across Tables 1-5, all results are single-run; no standard deviations or number of seeds are reported. This is particularly important for the headline comparison in Table 1, where the margin over SMURF is 1.11 mAP in the full area and 2.08 mAP in the corridor; these margins may be within run-to-run variation for 3D detectors. Please report mean +/- std over at least three seeds for the main comparisons and the ablations, or otherwise justify the stability of the improvements.","section":"Tables 1-5, experimental methodology"},{"comment":"Table 3's formatting makes the ablation hard to verify. Several rows appear to contain the same set of checkmarks under the printed column headers, and the text's reference to rows '(e) and (d)' does not match the reader's parsing of which components are active. The incremental claims (SSOD vs GT, one-adapter vs two-adapter FRFD, LRFD contribution) need a table with explicit per-row component indicators and the exact configurations used; otherwise the ablation does not support the decomposition claimed in the text.","section":"Table 3, ablation study"}],"minor_comments":[{"comment":"The abstract and main text should say 'percentage points' rather than '%' for mAP improvements: 10.38% and 5.12% are relative changes only if the denominator is specified; the reported numbers are absolute mAP differences.","section":"Abstract and Section 'Main Results'"},{"comment":"The author affiliation contains a typo: 'Institude' should be 'Institute'.","section":"Author affiliations"},{"comment":"Equations (6)-(7): 'index function' should be 'indicator function'.","section":"Equations (6)-(7)"},{"comment":"The notation in Eq. (13) could be clarified: D_T is introduced as a set of selected targets, but then used as a mask in Eq. (14); please define it as a binary mask explicitly.","section":"Eq. (13)-(14)"},{"comment":"In Table 4, the row labeled SCKD uses a teacher trained on 1/4 labeled data and a student on 1/4 unlabeled data, whereas SCKD+ uses full unlabeled data; the text should state this explicitly to avoid confusion with the main SCKD configuration.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"No concerns beyond those stated in the report. The manuscript is a reasonable fit for a computer-vision conference; the requested sensitivity analysis and multi-run reporting are feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid within-subfield paper that ships code and shows a real gain for radar-only 3D detection. The new piece is the combination of a LiDAR-radar fusion teacher with semi-supervised output distillation, so the radar-only student never sees ground truth. That is a genuine extension of RadarDistill and BEVDistill, and the two-adapter feature distillation is a reasonable, if minor, technical detail.\n\nWhat the paper does well: the experiments are straightforward and the main claim holds. On VoD, the student reaches 52.08 mAP, +10.38 over the SECOND baseline and +1.11 over SMURF, at 39.3 FPS. On ZJUODset, adding unlabeled data gives another clear boost, which supports the semi-supervised story. The ablations are reasonably complete: each component contributes, and Table 5 shows the fusion teacher beats a LiDAR-only teacher. Code is available.\n\nSoft spots, in order of importance. First, the SSOD pseudo-label threshold sigma is set to 0.1 and is never ablated. The stress-test note is fair here: if the final mAP is sensitive to that threshold, the 'SSOD is better than GT' claim loses force. It is not a fatal flaw — the teacher is strong and 0.1 may be fine — but a sigma sweep is the obvious missing experiment. Second, all tables report single runs without error bars; a 1.11 mAP margin over SMURF could be within run-to-run noise, especially since the published numbers are not re-run. Third, the ablation table formatting is confusing: row labels and column groupings are easy to misread. Minor.\n\nOverall: the paper is worth a serious referee. My recommendation: conditionally accept, requiring a sigma sensitivity analysis and ideally multiple runs with variance. The central mechanism is sensible and the reported improvement over the baseline is large enough that a few missing robustness checks should not sink it.","headline":"Solid radar-only detection paper with a genuinely semi-supervised cross-modal distillation setup; the unablated confidence threshold is the main thing to push on in review.","tokens_in":13740,"tokens_out":2370,"would_cite":true,"duration_ms":18829,"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":"Radar-only detector gains 10.4 mAP points via knowledge distillation from a lidar-radar teacher.","keywords":["4D radar","3D object detection","knowledge distillation","semi-supervised learning","cross-modality","autonomous driving","pseudo-labeling"],"falsifier":"Train the same SCKD pipeline on a dataset where the teacher is deliberately degraded, such as a subset of scenes with heavy occlusions or sensor noise, and measure whether the student's accuracy tracks the teacher's errors. If the student fails to improve over baseline or reproduces the teacher's systematic mistakes, the SSOD pseudo-label supervision is the point of failure.","tokens_in":12804,"feed_emoji":"📡","tokens_out":4160,"duration_ms":31480,"temperature":0.7,"pith_summary":"The paper tries to show that a cheap, radar-only 3D object detector can be made nearly as accurate as expensive multi-sensor systems by distilling knowledge from a lidar-radar fusion teacher network. The central claim is that semi-supervised distillation from a bi-modality teacher, using the teacher's own confident outputs as pseudo-labels, transfers enough feature quality to a radar-only student to boost its mean average precision by 10.38 points over the same network trained with ground truth. If true, this matters because autonomous vehicles could rely on 4D radar alone for perception, keeping real-time speed and all-weather robustness without the cost and latency of lidar or camera fusion.","feed_headline":"Radar-only detector gains 10.4 mAP points via distillation","feed_subtitle":"Semi-supervised pipeline transfers lidar-radar fusion knowledge to a real-time radar student.","key_machinery":"The load-bearing mechanism is the combination of three distillation signals. An adaptive fusion module in the teacher produces a fused lidar-radar feature map by learning per-modality weights, with a dropout gate that randomly discards one modality during training. Two feature-distillation losses, LRFD (lidar-to-radar) and FRFD (fusion-to-radar), align the student's radar features to the teacher's lidar and fused features using adapter convolutions. A semi-supervised output distillation loss treats the teacher's thresholded predictions as pseudo-labels, replacing ground truth supervision entirely.","core_discovery":"The paper proposes SCKD, a distillation framework in which a lidar-radar fusion teacher (a SECOND backbone with an adaptive fusion module that weights and concatenates the two modalities, plus random modality dropout) supervises a radar-only SECOND student. Two feature-distillation losses, LRFD and FRFD, map the student's radar feature into the teacher's lidar and fused feature spaces through learned adapters, and a semi-supervised output distillation loss uses the teacher's detections with confidence above 0.1 as pseudo-labels. The student is trained entirely on teacher supervision, with no ground-truth labels, and at inference only the student runs. On the View-of-Delft dataset the student reaches 52.08 mAP versus 41.70 for the baseline SECOND, surpassing all prior radar-only methods and all radar-camera fusion methods except LXL, while running at 39.3 FPS. On ZJUODset, adding unlabeled data improves the moderate-level mAP by 5.12 points over the same baseline.","pith_inferences":["If SCKD's gains hold across sensors and weather conditions, radar-based perception could become a primary modality for low-cost autonomy, since the distillation pipeline transfers semantic richness without adding sensor cost at deployment.","The success of SSOD suggests that a well-trained teacher's noisy but high-recall outputs at a low confidence threshold may serve as a more effective training signal than ground-truth boxes for the student, because teacher false positives carry contextual information that manual labels omit.","A testable extension is to apply SCKD to other backbone architectures and radar datasets, such as TJ4DRadSet or RADIATE, to check whether the gain is specific to SECOND and VoD or generalizes across architectures and domains.","A possible failure mode to probe is distribution shift in unlabeled data: if the unlabeled pool comes from different weather or traffic density than the teacher's training set, the pseudo-labels may become systematically biased, and student performance on labeled validation splits would reveal this."],"forward_implications":["A radar-only detector can reach or exceed the accuracy of most lidar-camera or radar-camera fusion detectors while running an order of magnitude faster.","The student network can be trained without any labeled data beyond what the teacher needs, so large unlabeled corpora such as raw driving logs become usable for improving radar detection.","Feature-level distillation from a bi-modality teacher is more effective than from a single-modality teacher because the fused feature space is closer to the student's radar-only space.","Using two separate adapters in FRFD rather than a single upscaling convolution preserves the fused feature structure and improves transfer."],"supporting_citations":[{"why":"SECOND is the baseline detector and the backbone for both teacher and student; its numbers are the direct comparison point for the reported mAP gains.","marker":"(Yan, Mao, and Li 2018)"},{"why":"The View-of-Delft dataset is the primary evaluation benchmark, providing the lidar, radar, and camera data used in all main experiments.","marker":"(Palffy et al. 2022)"},{"why":"The ZJUODset dataset is the second evaluation benchmark, used to demonstrate improvements from additional unlabeled data.","marker":"(Xu et al. 2023)"},{"why":"RadarDistill is the prior radar-only distillation method that SCKD builds on and compares against; its limitations motivate the bi-modality teacher design.","marker":"(Bang et al. 2024)"},{"why":"SMURF is the previous state-of-the-art radar-only detector on VoD, serving as the main radar-only accuracy benchmark that SCKD surpasses.","marker":"(Liu et al. 2023)"},{"why":"LXL is the best radar-camera fusion method on VoD, the key comparison for the claim that SCKD outperforms most fusion methods at much higher speed.","marker":"(Xiong et al. 2023)"}],"fun_headline_variants":["Radar-only 3D detection gains 10.4 mAP via cross-modal distillation","Semi-supervised distillation transfers lidar knowledge to radar-only model","Radar student learns from lidar fusion teacher, boosts mAP by 10.4","No labels needed: radar detector improves 10.4 mAP via distillation","Cross-modal knowledge distillation lifts radar-only detection 10.4 mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The teacher's detection outputs, kept as pseudo-labels when their confidence exceeds 0.1, are good enough substitutes for human-annotated ground truth; if those pseudo-labels are too noisy, the student will absorb the teacher's errors.","fun_headline_variants_meta":{"raw":{"variants":["Radar-only 3D detection gains 10.4 mAP via cross-modal distillation","Semi-supervised distillation transfers lidar knowledge to radar-only model","Radar student learns from lidar fusion teacher, boosts mAP by 10.4","No labels needed: radar detector improves 10.4 mAP via distillation","Cross-modal knowledge distillation lifts radar-only detection 10.4 mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000717,"raw_usage":{"total_tokens":3261,"prompt_tokens":1021,"completion_tokens":2240,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":2136}},"tokens_in":637,"tokens_out":2240,"duration_ms":13614,"temperature":1.0,"reasoning_tokens":2136,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:05:46.913846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same SCKD pipeline on a dataset where the teacher is deliberately degraded, such as a subset of scenes with heavy occlusions or sensor noise, and measure whether the student's accuracy tracks the teacher's errors. If the student fails to improve over baseline or reproduces the teacher's systematic mistakes, the SSOD pseudo-label supervision is the point of failure.","supporting_citations":[{"cited_title":"F.; and Gavrila, D","cited_arxiv_id":null,"evidence_quote":"The View-of-Delft dataset is the primary evaluation benchmark, providing the lidar, radar, and camera data used in all main experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RadarDistill is the prior radar-only distillation method that SCKD builds on and compares against; its limitations motivate the bi-modality teacher design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SMURF is the previous state-of-the-art radar-only detector on VoD, serving as the main radar-only accuracy benchmark that SCKD surpasses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LXL is the best radar-camera fusion method on VoD, the key comparison for the claim that SCKD outperforms most fusion methods at much higher speed."}],"review_version":1}