{"id":"14d578f8-5a33-481b-93cb-a35d26bed80b","arxiv_id":"2501.18162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"IROAM improves roadside monocular 3D object detection by jointly training on vehicle-side and roadside images and aligning only semantic query features across the two views.","lead":"IROAM is a training method that lets a roadside camera detector learn from vehicle camera data by separating each detected object's semantic features from its geometric features and applying contrastive learning only to the semantic part. Tests on two vehicle-infrastructure cooperative datasets show higher 3D detection accuracy on roadside images than training on roadside data alone or on a simple mix of both datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paired-data requirement is the weak link: Eq. 7 needs synchronized roadside and vehicle images with depth ground truth in both domains, so the method cannot exploit the abundant unpaired AV datasets cited as motivation, and this setting is never tested.","rationale":"The reader's weakest_assumption identifies exactly the condition that is most load-bearing for the paper's motivating claim: IROAM requires paired, synchronized vehicle-side and roadside images with ground-truth depth in both domains. My reading agrees and sharpens why this matters. The paper's stated motivation is to leverage the abundance of existing vehicle-side datasets, but the method as specified in Eq. 7 cannot train on KITTI, nuScenes, or Waymo because those datasets have no paired roadside images and, in most cases, no per-frame depth supervision from a second viewpoint. The experiments all use DAIR-V2X-V and DAIR-Seq-V, which are cooperative datasets with synchronized roadside and vehicle frames, so the core scenario of 'abundant unpaired AV data' is never tested. The internal logic of the method is coherent on paired data, and the ablation study gives some evidence that the contrastive and decoupling components contribute beyond simply mixing data. However, the applicability of the method to the motivating data regime is a separate empirical question, and the paper provides no evidence on it. I do not recommend changing the reader's CONDITIONAL verdict: the reported numbers can stand as a result on paired VIC data, but the paper should either demonstrate the unpaired/depth-free setting works or explicitly restrict its claims to the paired setting. The proposed concrete test would settle whether the synchronized-pairing requirement is truly essential or whether the method is more general than the current formulation implies.","tokens_in":11204,"tokens_out":10116,"duration_ms":104163,"concrete_test":"Run the Table I DAIR-V2X-V to DAIR-V2X-I experiment again, but for the vehicle branch use temporally unpaired DAIR-V2X-V frames (e.g., randomly permuted timestamps) while keeping the same marginal class and depth distribution. If the AP3D gap relative to Addon disappears or shrinks below about 1 AP, synchronized pairing is essential and IROAM cannot deliver its stated benefit from the unpaired AV datasets that motivated it. A second arm should disable vehicle-side depth supervision (drop L_dmap,v and the depth terms in L_pair,v) to test whether depth-free vehicle data can be exploited at all.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IROAM lets abundant vehicle-side data improve roadside monocular 3D detection. The training objective in Eq. 7 makes this conditional on a specific data regime: every update consumes a paired roadside image and vehicle-side image, and both branches need ground-truth depth maps (L_dmap,r, L_dmap,v) and per-object labels (L_pair,r, L_pair,v) for Hungarian matching and depth supervision. DAIR-V2X happens to provide synchronized vehicle-infrastructure frames with LiDAR-derived depth, which is exactly why the experiments are possible. But the datasets the paper cites as the motivation—KITTI, nuScenes, Waymo—are vehicle-only and have no paired roadside images, so IROAM cannot be trained on them at all. In other words, the paper does not demonstrate learning from the abundant AV data domain; it demonstrates learning from an already-paired, already-depth-labeled VIC dataset. This is a mismatch between the motivating problem and the method, and it is load-bearing because the value proposition of the paper is that roadside annotation costs can be amortized by reusing existing AV data. No experiment in Table I addresses this setting: all IROAM rows use DAIR-V2X-V or DAIR-Seq-V, which are paired with the roadside evaluation split and contain depth supervision. The concern is not internal inconsistency—the method works as described on paired data—but the claimed applicability to “tremendous and diverse vehicle-side data” is unverified and likely false for most such data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IROAM, a framework for roadside monocular 3D object detection that jointly trains a roadside branch and a vehicle-side branch on top of MonoDETR. It introduces an In-Domain Query Interaction module that produces object queries from content and depth features, and a Cross-Domain Query Enhancement module that decouples queries into semantic and geometry parts and applies contrastive learning only to the semantic parts. Experiments on DAIR-V2X and V2X-Seq report consistent improvements over roadside-only and simple data-merging baselines on AP3D and APBEV at IoU thresholds of 0.7 and 0.5, including cross-dataset vehicle-side training and an analysis of data imbalance.","tokens_in":11588,"tokens_out":6162,"duration_ms":60766,"significance":"If the reported results are reproducible, the paper demonstrates a useful technique for transferring vehicle-side monocular representations to roadside detection, and the modular design plus cross-dataset experiments (DAIR-V2X and V2X-Seq) are strengths. The analysis of data imbalance in Figure 5 is a valuable addition. However, the central claim about learning from abundant vehicle-side data is not fully supported because the method as described requires paired, depth-annotated vehicle-roadside images; the claimed applicability to vehicle-only datasets such as KITTI, nuScenes, and Waymo is therefore unverified. The paper would be strengthened by a precise pairing protocol, a more standard contrastive formulation, and either experiments on unpaired data or a more careful statement of scope.","major_comments":[{"comment":"The overall loss in Eq. (7) requires, for every training step, a synchronized vehicle-roadside image pair with ground-truth depth maps and object labels in both domains. However, the Introduction and Conclusion motivate the method by \"tremendous and diverse vehicle-side data\" from vehicle-only datasets such as KITTI, nuScenes, and Waymo. These datasets do not contain synchronized roadside images or roadside ground truth, so IROAM cannot be trained on them as described. All experiments in Table I use DAIR-V2X-V or DAIR-Seq-V as vehicle-side data, both of which are paired with the evaluation roadside dataset. The paper does not test an alternative regime, such as pretraining the vehicle branch on unpaired vehicle data and fine-tuning on sparse vehicle-roadside pairs. The central claim that IROAM learns from the abundant autonomous-vehicle data domain is therefore not supported by the reported experiments.","section":"III-D"},{"comment":"The paper does not specify how the vehicle-side and roadside images are paired. This matters because the contrastive loss in Eq. (6) treats every query matched to any ground-truth object in the merged positive set QP as mutually similar (Eq. (5), label 1 for all i,j in P). If images are randomly paired or come from different datasets (as in Table I rows that use DAIR-Seq-V with DAIR-V2X-I), then positive pairs do not correspond to the same objects or even the same scenes, and the semantic-alignment rationale illustrated in Figure 1 is not realized. Please specify the pairing procedure, whether synchronized frames from the VIC dataset are used, and whether positive labels are based on object correspondence or merely on 'any matched object'.","section":"III-C"},{"comment":"In Eqs. (4)-(6), the similarity is defined as a sigmoid of the cosine similarity, so the maximum possible value of s_i,j is about 0.731 while the target label is 1 in Eq. (5). The L1 loss in Eq. (6) therefore never reaches zero for positive pairs, and the gradient behavior of this formulation is not analyzed. Additionally, the positive set QP merges queries from both domains without considering object category or identity, which may encourage alignment of different object types. Please justify the sigmoid-plus-L1 formulation or replace it with a standard InfoNCE or cosine-embedding loss, and discuss the effect of cross-category positives.","section":"III-C"},{"comment":"In Table III, the row labeled with only \"Veh Data\" (CL and DC absent) reports AP3D easy of 24.61, whereas the Addon baseline in Table I, which also trains on vehicle-side and roadside data together, reports 26.74. The text states that \"Veh Data means to train the model with both roadside and vehicle data,\" which should be equivalent to Addon. The discrepancy is unexplained and makes it difficult to attribute the gains in the ablation to the individual components. Please clarify the experimental setup of Table III (e.g., number of vehicle images per epoch, use of pair sampling, and any differences from the Addon configuration in Table I) and reconcile the two tables.","section":"IV-C"}],"minor_comments":[{"comment":"Eq. (6) contains a typo: \"lable\" should be \"label\".","section":"III-C"},{"comment":"All experiments are reported as single runs on the validation split; please provide multiple seeds or at least a statement of variance, since the reported improvements are moderate in some cells.","section":"IV-A"},{"comment":"The caption of Figure 2 is dense and does not introduce the notation f^e_{C,*} and f^e_{D,*}; please define these in the caption or point explicitly to Section III-A.","section":"Figure 2"},{"comment":"Equations (2) and (3) use overlapping weight notation: Eq. (2) uses lambda_1..4 for matching cost while Eq. (3) adds lambda_5..7 for the final pair loss. Consider clarifying that these are two distinct weighted sums and that the same lambda symbols appear in both.","section":"III-B"},{"comment":"The first sentence of the Introduction contains a capitalization error: \"In autonomous driving, The perception...\" should be \"the perception...\".","section":"I"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution, but the scope of the claim is overstated relative to the experiments. For a journal, I would expect either (i) an experiment that actually uses unpaired vehicle-only data (e.g., pretraining on KITTI and fine-tuning on DAIR-V2X-I), or (ii) an explicit limitation statement that the method requires VIC-style synchronized pairs. The revision should also resolve the ablation/Addon discrepancy and clarify the pairing protocol. The novelty relative to MonoDETR is modest but acceptable if the contributions are validated. The single-run, val-only evaluation is a concern but not blocking if the authors add error bars or multiple seeds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The method is a sensible extension of the MonoDETR line, and the reported gains on roadside monocular 3D detection are consistent. But be careful with the framing: the paper sells itself as exploiting abundant vehicle-side data while the experiments only use vehicle-side data from vehicle-infrastructure cooperative datasets. That is a weaker claim than advertised.\n\nThe new trick is to split each DETR query into a semantic half and a geometry half, then run contrastive learning only on the semantic half. That is a small but clean idea, and the ablation shows it matters: adding the decoupling raises AP3D(IoU=0.7) moderate from 15.73 to 16.11, and the full method beats both roadside-only and the naive Addon baseline on every metric. The imbalance analysis (Fig. 5) is also a nice touch; it shows IROAM is more compute-efficient than Addon when vehicle-side data dominates.\n\nThree weaknesses. First, no experiment on generic AV data. KITTI, nuScenes, and Waymo are all named as motivation, but all training uses paired VIC vehicle-side data with depth supervision. The method does not need temporally synchronized vehicle/roadside frames—it just needs one road and one vehicle image per batch, and depth labels are available in KITTI/nuScenes/Waymo from LiDAR. So the stress-test note overstates the pairing requirement, but the underlying point stands: the open question is whether IROAM transfers to truly diverse, non-VIC vehicle data. An experiment with KITTI as vehicle side would settle it.\n\nSecond, the contrastive objective in Eq. 5 labels every pair of positive queries as a positive, regardless of whether the objects are the same category. The model pulls all objects in an image toward one another, which seems likely to wash out object-class distinctions. The L1 loss on a sigmoid cosine similarity is also under-justified. An ablation with a class-aware contrastive loss, or at least a discussion, is needed.\n\nThird, all numbers are single-run on the val split. Gains of 4-5 AP are large, but no error bars means I can't tell if the edge over Addon is meaningful. The paper should rerun with a few seeds. Also, no code release, which makes these issues harder to check.\n\nBottom line: I'd send it to review. The core idea is publishable and the experiments, though narrow, are internally coherent. The authors need to broaden the evaluation, tighten the contrastive loss, and report variance.","headline":"A clean but narrow contrastive-learning trick for roadside 3D detection; the strong 'abundant AV data' claim is not yet backed by the experiments.","tokens_in":12050,"tokens_out":4283,"would_cite":true,"duration_ms":43523,"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":"IROAM improves roadside monocular 3D detection by aligning only semantic object queries across vehicle-side and roadside views, lifting AP3D at IoU 0.7 from 24.31 to 28.89 on the easy split.","keywords":["roadside monocular 3D object detection","contrastive learning","semantic-geometry decoupling","cross-domain learning","vehicle-infrastructure cooperation","object queries","depth-guided transformer"],"falsifier":"Rerun IROAM's training on the same benchmark but with the vehicle-side and roadside frames randomly re-paired within each batch; if the roadside AP3D improvement over the data-merging baseline remains, then the claimed benefit does not depend on real cross-view correspondences, contradicting the method's design.","tokens_in":11036,"feed_emoji":"🚗","tokens_out":8910,"duration_ms":79772,"temperature":0.7,"pith_summary":"This paper argues that roadside monocular 3D object detectors can be made more accurate by training them jointly on abundant vehicle-mounted camera data, despite a substantial viewpoint gap between the two domains. The central idea is to extract object queries from both domains and align only their semantic parts across views, leaving geometric parts such as depth and orientation to be supervised separately within each domain. On the DAIR-V2X vehicle-infrastructure benchmark, the proposed IROAM method raises roadside AP3D at IoU 0.7 from 24.31 to 28.89 on the easy split, and from 13.54 to 16.11 on the moderate split. The gains hold when the vehicle-side dataset is swapped, suggesting cross-domain transfer rather than overfitting.","feed_headline":"Roadside 3D detection gains 19% from vehicle-camera data","feed_subtitle":"Aligning only semantic object queries lets roadside detectors borrow knowledge from abundant vehicle-side images.","key_machinery":"The load-bearing mechanism is semantic-geometry decoupled contrastive learning over transformer object queries. Each query produced by a depth-guided detector is split into semantic and geometry halves; cosine similarity between semantic halves of positive cross-domain pairs is pulled toward 1, negative pairs toward 0, while geometry halves are never compared across domains. A Query Sampler selects positive and negative samples using Hungarian matching scores in each domain. This lets the model borrow category and appearance knowledge from the vehicle domain without forcing depth and geometry distributions to match.","core_discovery":"The paper's discovery is that the reason vehicle-side data has not helped roadside detectors is not a lack of useful information but a geometry mismatch: the same object appears with very different depth distributions from the two viewpoints. IROAM therefore decouples each object query into a semantic half and a geometry half, runs contrastive learning only on the semantic half, and keeps geometry supervised per domain. This yields roadside AP3D(IoU=0.7) of 28.89/16.11/15.85 on easy/moderate/hard, compared with 24.31/13.54/13.17 for the roadside-only baseline and 26.74/15.36/13.84 for naive data merging. The same training also improves vehicle-side detection, and the benefit survives replacing the vehicle-side dataset, supporting the claim that semantic alignment transfers across vehicle and roadside domains.","pith_inferences":["If semantic-geometry decoupling is the active ingredient, the same recipe may transfer to other viewpoint gaps, such as drone-to-ground cameras or different camera heights, where semantics match but geometry does not.","Because the contrastive loss operates on object queries rather than pixels, it could compose with any query-based detector; a natural test is swapping the depth-guided backbone for a lighter query-based detector.","A testable extension is to relax the paired-synchronization requirement using pseudo-pairs or self-supervised depth, which would let the method exploit large unpaired vehicle-side collections.","The imbalance experiment suggests a fixed pool of vehicle-side data can serve many roadside locations, pointing toward pretraining a roadside detector once on a large vehicle-side corpus."],"forward_implications":["Roadside perception for vehicle-infrastructure cooperative driving can be improved by reusing existing vehicle-side datasets, reducing the need for costly roadside annotation.","Cross-domain training also improves vehicle-side detection, so the benefit is bidirectional rather than one-way transfer.","The improvement persists when the vehicle-side dataset is changed, indicating the learned semantic alignment generalizes across datasets.","The method remains useful when roadside data is scarce relative to vehicle-side data, with no extra inference cost.","Decoupling semantic from geometry contributes roughly 1.99 AP3D over contrastive learning on full queries, isolating the mechanism that drives the gain."],"supporting_citations":[{"why":"Supplies the DAIR-V2X vehicle-infrastructure dataset of paired roadside and vehicle-side frames used for the main experiments and comparisons.","marker":"[1]"},{"why":"Provides the depth-guided transformer detector and depth prediction losses that IROAM builds on for object-query extraction.","marker":"[15]"},{"why":"Defines the easy/moderate/hard difficulty splits and the AP3D/APBEV evaluation protocol used to report results.","marker":"[5]"},{"why":"Demonstrates contrastive learning across vehicle-side and roadside point clouds, motivating cross-view contrastive alignment for 3D detection.","marker":"[25]"},{"why":"Supplies the V2X-Seq dataset used to test whether IROAM's improvement survives swapping the vehicle-side data source.","marker":"[2]"},{"why":"Defines the Hungarian matching procedure used to match object queries to ground-truth labels and to select positive and negative samples for contrastive learning.","marker":"[13]"}],"fun_headline_variants":["Roadside 3D detection: semantic alignment unlocks vehicle data","IROAM: Semantic-only contrast transfers vehicle knowledge to roadside","Roadside 3D detectors borrow wisdom from vehicle views via semantic alignment","Vehicle data lifts roadside 3D detection by aligning only object semantics","19% roadside 3D gain from aligning semantics not geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method needs paired, synchronized vehicle-side and roadside images with ground-truth depth in both domains during training, so it does not apply as described to the common case of unpaired or depth-unlabeled vehicle-side data.","fun_headline_variants_meta":{"raw":{"variants":["Roadside 3D detection: semantic alignment unlocks vehicle data","IROAM: Semantic-only contrast transfers vehicle knowledge to roadside","Roadside 3D detectors borrow wisdom from vehicle views via semantic alignment","Vehicle data lifts roadside 3D detection by aligning only object semantics","19% roadside 3D gain from aligning semantics not geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3663,"prompt_tokens":896,"completion_tokens":2767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2678}},"tokens_in":512,"tokens_out":2767,"duration_ms":19242,"temperature":1.0,"reasoning_tokens":2678,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:26:42.705480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun IROAM's training on the same benchmark but with the vehicle-side and roadside frames randomly re-paired within each batch; if the roadside AP3D improvement over the data-merging baseline remains, then the claimed benefit does not depend on real cross-view correspondences, contradicting the method's design.","supporting_citations":[{"cited_title":"Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the DAIR-V2X vehicle-infrastructure dataset of paired roadside and vehicle-side frames used for the main experiments and comparisons."},{"cited_title":"Monodetr: Depth-guided transformer for monocular 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the depth-guided transformer detector and depth prediction losses that IROAM builds on for object-query extraction."}],"review_version":1}