acceptodds
Under review as a conference paper at ICLR 2027

EviScope: Evidence-Conditioned Sampling for Long-Range 3D Detection

Abstract

As autonomous driving extends to longer ranges, reliable 3D detection depends on using multiple sensing modalities well. Camera–radar fusion, pairing visual semantics with radar's reach, is an important direction. Query-based camera–radar detectors sample features around query positions. On MAN TruckScenes, recall varies substantially across object-level radar-return counts within the same range band. This motivates testing local radar evidence as an additional conditioning signal. We introduce , evidence-conditioned query sampling: each query adapts its spatial extent and temporal weights to local evidence. The current frame's returns form a non-learned bird's-eye evidence field, which each query reads to predict a sampling radius and frame weights. It adds parameters, of the detector, with no latency increase in our timing protocol. The module-free recipe exceeds the published mAP values in our comparison. At eight epochs, single runs reach / mAP / NDS without the module and / with it. Across seven matched four-epoch pairs, the module changes NDS by and mAP by (mean SE), leaving its incremental benefit unresolved. An exploratory subset gain fails an internally pre-specified test on two further seeds. The sampling radius tracks local evidence in all four probed runs, yet a large penalty from replacing temporal weights also occurs in a control lacking query-local evidence. These results distinguish learned allocation and inference-time dependence from demonstrated detection gains.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D object detection, camera-radar fusion, long-range detection, query-based detectors, MAN TruckScenes, evidence-conditioned sampling, temporal fusion

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_kiSaOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

EviScope studies whether local radar evidence should condition spatial and temporal query sampling for long-range camera-radar 3D detection. The method constructs a parameter-free BEV evidence field and uses a 5,450-parameter shared conditioner to predict a radial sampling scale and temporal weights. On MAN TruckScenes, the authors first establish a strong RaCFormer training recipe and then evaluate the module with seven matched four-epoch pairs, one eight-epoch pair, allocation probes, substitution tests, and an internally pre-specified held-out seed test. The recipe is competitive, and the module reliably learns evidence-associated allocation, but its incremental detection benefit remains statistically unresolved (+0.31 +/- 0.36 NDS and -0.27 +/- 0.22 mAP over seven pairs).

Critical strengths
  1. The empirical methodology is unusually careful. Matched seeds, uncertainty estimates, a held-out confirmation test, and an evidence-blind control prevent the paper from overstating attractive but non-replicating findings. This is consequential because it turns an inconclusive module result into a credible lesson about evaluation in a noisy 3D-detection regime.
  2. The proposed conditioner is technically lightweight and well motivated, and the supplementary package is detailed enough to inspect the key implementation choices and reproduce the analyses. The strong module-free recipe is itself practically useful.
Critical weaknesses
  1. The central architectural claim is not supported by a reproducible end-to-end gain: the paired mAP estimate is negative and the NDS estimate is smaller than its standard error. Therefore acceptance rests mainly on the diagnostic methodology and negative result, not on EviScope as an improved detector.
  2. The evidence-specific mechanism is not isolated from learned geometry/temporal conditioning under the full schedule. The paper identifies the needed geometry-only matched-seed control but does not report it, leaving the causal value of query-local radar evidence open.
Questions for authors
  1. Can the authors provide full-schedule matched-seed results for a geometry-only or evidence-shuffled conditioner with the same capacity and learned temporal branch? This would materially determine whether local radar evidence, rather than added adaptive capacity, is useful.
  2. How sensitive are the paired conclusions to checkpoint selection and to evaluation on a second split or dataset? Even a limited cross-dataset check would affect how broadly the negative/diagnostic conclusion should be interpreted.
2 comments
AuthorsOpenReview ↗

Response to Reviewer kiSa

Thank you for identifying the full-schedule controls that can test the contribution of local radar evidence. We report geometry-only controls at three seeds, shuffled-evidence controls at two seeds, and the sensitivity analyses below. All ± values are mean ± SE over matched seeds.

W1. Reproducible gain over the module-free detector. The original seven four-epoch pairs remain +0.31 ± 0.36 NDS and −0.27 ± 0.22 mAP. Our new evidence distinguishes the value of the input within the conditioner from the module's total gain:

  • Geometry-only control (Table 26, seeds 0–2). Its observed mean NDS is 38.14 versus the baseline's 38.24; its mAP is lower at each seed (−0.49, −1.01, −0.86). Adding the evidence input gives +0.87 ± 0.21 mAP (+1.12, +1.04, +0.46) and +1.25 ± 0.72 NDS (+1.62, +2.26, −0.14). Translation error improves in all three pairs: 0.675→0.636, 0.657→0.624, 0.662→0.648. Same-node re-evaluation gives +0.91 ± 0.22 mAP and +1.26 ± 0.73 NDS (Table 26†).
  • Eight-epoch baseline pairs (Table 26). Seed 1 gives 40.08 versus 38.89 NDS (+1.19), with mAP 26.49 versus 26.62 (−0.13); both arms were trained and evaluated on the same node. Together with seed 0 (+0.67 NDS, +0.26 mAP), the two pairs average +0.93 ± 0.26 NDS and +0.07 ± 0.20 mAP. Both completed eight-epoch pairs therefore improve NDS at matched training budgets.

W2 and Q1. Full-schedule geometry-only and shuffled-evidence controls.

  • Geometry-only, seeds 0–2: The same 5,450-parameter module receives zeroed evidence channels while retaining query geometry, radial/temporal heads, identity initialisation, and the auxiliary target from true counts. Only access to the five evidence inputs changes. Figure 7 shows inverse evidence/radius association in EviScope within every range band (Spearman −0.89 to −0.95, five seeds), versus nearly constant 0.600–0.607 radii in geometry-only.
  • Shuffled evidence, seeds 0–1: Evidence vectors are permuted across the 900 queries of each sample at each training/evaluation call. The controls score 25.19/37.30 and 26.31/38.25 (mAP/NDS). EviScope − shuffled is +1.28/+2.89 and +0.20/+1.99, averaging +0.74 ± 0.54 mAP and +2.44 ± 0.45 NDS. The permutation preserves each frame's evidence distribution while disrupting alignment. These results support aligned conditioning over the permutation control.

Q2. Checkpoint sensitivity and another split/dataset.

  • Checkpoint protocol (App. T): Main comparisons use the final checkpoint of a fixed schedule. Re-evaluating epochs 2, 3 and 4 on the control's node gives across-seed mean differences of +0.11 ± 0.46 / +0.20 ± 0.76, +0.33 ± 0.15 / +0.55 ± 0.50, and +0.91 ± 0.22 / +1.26 ± 0.73 (mAP/NDS). The seed-mean advantage increases over these checkpoints; individual seed trajectories vary (App. T). The main results use the scheduled final checkpoint, without selecting the best validation epoch.
  • Scene conditions (Table 27): Mean ΔNDS is positive in 19/20 official tags with at least 300 frames (rural: −0.11); all three seeds are positive in 7/20. These are overlapping validation subsets, not independent splits or datasets.
  • Second split/dataset: These checks use TruckScenes validation. We do not report a second-dataset evaluation.
AuthorsOpenReview ↗

Follow-up to Reviewer kiSa: Completed Three-Seed Controls (W1, Q1)

Thank you again for your careful review. Since our initial reply to your review—which reported two seeds each for the shuffled-evidence control (W2/Q1) and the eight-epoch baseline comparison (W1)—seed 2 has now completed for both experiments (with the shuffled-evidence update first noted in our 15 November comment, and the third eight-epoch pair just completed on the same node):

Matched comparisonCompleted seedsEviScope − reference (mAP)EviScope − reference (NDS)
Geometry-only conditioner (4 ep, W2/Q1)0–2 (3/3)+0.87 ± 0.21 (+1.12, +1.04, +0.46)+1.25 ± 0.72 (+1.62, +2.26, −0.14)
Shuffled-evidence conditioner (4 ep, W2/Q1)0–2 (3/3)+0.86 ± 0.34 (+1.28, +0.20, +1.11)+1.75 ± 0.74 (+2.89, +1.99, +0.36)
Module-free baseline at 8 epochs (W1)0–2 (3/3)+0.13 ± 0.13 (+0.26, −0.13, +0.27)+0.66 ± 0.31 (+0.67, +1.19, +0.12)

Values are mean ± SE in percentage points:

  • Causal isolation within the conditioner (W2/Q1): Both four-epoch controls retain the exact 5,450 parameters, learned radial and temporal heads, identity initialisation, and auxiliary target. At seed 2, the shuffled control scores 24.66 / 37.37 against EviScope's 25.77 / 37.73 (+1.11 mAP, +0.36 NDS), so both mAP and NDS improve in all three shuffled pairs, and all three shuffled runs also fall below the module-free baseline on both metrics (−0.65 / −1.18, −0.17 / −0.60, −1.51 / −0.02).
  • Matched baseline comparisons (W1): At seed 2 (8 epochs, both arms trained and evaluated on one node), EviScope reaches 26.65 / 39.27 against the baseline's 26.38 / 39.15 (+0.27 mAP, +0.12 NDS). Across all three eight-epoch pairs, NDS is positive in every pair (+0.66 ± 0.31), alongside the seven four-epoch pairs (+0.31 ± 0.36 NDS, −0.27 ± 0.22 mAP; Table 7).

Could you please let us know whether these completed three-seed controls address your W2/Q1, and which remaining concerns—whether the four-epoch end-to-end margin (W1), single-benchmark scope (Q2), or points raised elsewhere in the discussion—weigh most heavily in your current assessment? We would greatly appreciate the opportunity to address them directly.

Reviewer_qhcnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EviScope, which uses local radar evidence to modulate sampling extent and temporal fusion within RaCFormer's polar-query and adaptive ray-sampling framework. The method constructs a five-channel BEV evidence field from current-frame radar returns, extracts local statistics at each predicted query center, and combines them with query geometry to predict a sampling-radius scale and fusion weights over eight frames. During training, an Evidence Head supervises the shared MLP through an auxiliary density loss; this branch is not used at inference. The added module contains approximately 5,450 parameters.

Across seven matched-seed comparisons on the MAN TruckScenes validation set, the paired changes in NDS and mAP are +0.31±0.36+0.31\pm0.36 and −0.27±0.22-0.27\pm0.22, respectively (mean ±\pm standard error, Table 7). The gain on the distant, low-evidence subset did not pass the subsequent replication test on two held-out training seeds.

Critical strengths

S1. Thorough reporting of experimental uncertainty and unsuccessful results. The authors report both positive and negative differences across seven matched seeds, distinguish exploratory subset analyses from tests on held-out seeds, and disclose results that did not replicate. This helps avoid interpreting an incidental gain from a single run as a reliable improvement (Secs. 4.3–4.4, Table 7).

S2. Careful interpretation of attribution evidence. The authors observe that replacing temporal weights also causes a performance drop in a control that does not correctly read local evidence. They therefore do not equate inference-time dependence with a causal benefit from evidence conditioning (Sec. 5.2).

Critical weaknesses

W1. The independent contribution of local evidence conditioning remains unclear. Polar queries and ray sampling are inherited from RaCFormer, while prior work such as SparseBEV already uses query-guided adaptive spatiotemporal sampling. The key question is whether explicit local radar statistics provide benefits beyond query geometry or features. The existing geometry-only control is trained for only one epoch; a comparison using the same prediction heads and multiple matched seeds under the full training budget is still missing. The single-run evidence-blind experiment using incorrect sampling coordinates cannot adequately substitute for this control. In addition, the correspondence between radar-return counts inside ground-truth boxes and the evidence around predicted query centers has not been measured, leaving a gap between the motivating analysis and the actual conditioning input (Sec. 3.2, Appendix H).

W2. Reliable evidence for detection gains and the effectiveness of auxiliary supervision is lacking. The seven paired comparisons do not establish a stable detection improvement, and recall on the distant subset decreases by 0.67 and 0.59 percentage points on the two held-out seeds. The Evidence Head's auxiliary target is derived from the input counts rather than directly supervising an optimal sampling policy; better density prediction does not necessarily imply better detection. In Table 11, adding this branch increases NDS from 38.93 to 40.19, but mAP only increases from 26.46 to 26.47. In the single-epoch experiment in Table 12, adding the branch instead decreases NDS from 27.58 to 22.67. These experiments use different training schedules and are each based on a single run, so they do not establish that auxiliary supervision consistently improves the shared representation and detection performance. The radius- and temporal-weight-substitution diagnostics also do not isolate the independent benefit of local evidence (Tables 5–6). Furthermore, evaluation is limited to TruckScenes and the current detector; testing on another compatible camera–radar benchmark would help assess the scope of applicability.

Questions for authors
  1. How is the Evidence Head's auxiliary density loss weighted relative to the detection loss, and how is it aggregated across the six decoder iterations? Could the authors compare models with and without this branch across multiple seeds, keeping all other settings fixed, to assess whether the gain in Table 11 is reliable and explain the opposite result in Table 12?

  2. With the same learnable temporal branch, what are the paired differences between models trained from scratch with fixed and adaptive radii? What justifies predicting weights for historical frames from current-frame evidence alone?

  3. Method description and presentation. Section 3.2 is too brief. Even with Appendix K, evidence-field construction, query extent sqs_q, and the integration of radius scaling and temporal weights remain unclear. Please add equations or pseudocode for these operations and the training objective, and summarize the key ablation settings in one place.

6 comments
AuthorsOpenReview ↗

Response to Reviewer qhcn — Part 1/2: evidence controls and presentation

Thank you for the detailed review. To address the presentation concerns, we expanded §3.2 and Appendix K with Eqs. 1–3, Algorithm 1, a redrawn Figure 2 and an inherited/added component table (Table 15). Appendix U, Table 28 summarises experimental settings. This part addresses W1–W2; Part 2 addresses Q1–Q3. All ± values are mean ± SE across matched seeds.

W1. Contribution beyond query geometry/features and the object-count/query-evidence gap.

  • Method boundary: RaCFormer provides the polar queries, ray sampling and decoder; query-guided adaptive sampling also precedes our work. EviScope adds explicit, non-learned local radar statistics to a conditioner that modulates radial sampling and frame weights. It is a targeted addition, not a new detector backbone.
  • Full-schedule geometry-only control (App. T, Table 26): Seeds 0–2 run four epochs with the same 5,450 parameters, learned heads, identity initialisation and auxiliary target; only the five evidence inputs are zeroed. EviScope − geometry-only is +0.87 ± 0.21 mAP (+1.12,+1.04,+0.46) and +1.25 ± 0.72 NDS (+1.62,+2.26,−0.14). mATE falls in all three pairs (0.675→0.636, 0.657→0.624, 0.662→0.648). Same-node re-evaluation gives +0.91 ± 0.22 mAP and +1.26 ± 0.73 NDS. Geometry-only's observed mean NDS is 38.14 versus 38.24 for the baseline, with lower mAP in each pair (−0.49,−1.01,−0.86).
  • Shuffled-evidence control (Table 26): Permuting evidence across queries during training and evaluation preserves the per-frame distribution and breaks alignment. Seeds 0 and 1 score 25.19/37.30 and 26.31/38.25 (mAP/NDS), below EviScope by +1.28/+2.89 and +0.20/+1.99. The mean differences are +0.74 ± 0.54 mAP and +2.44 ± 0.45 NDS. This supports the contribution of query-aligned evidence.
  • Allocation probe (Figure 7): In all range bands and five EviScope seeds, radius decreases with evidence (Spearman −0.89 to −0.95); geometry-only stays near its lower bound (0.600–0.607, SD <0.001). The allocation pattern is reproduced across these seeds and distance bands.
  • Measured correspondence (App. S, Figure 5, Table 25): We sampled 300 validation frames and matched GT boxes within 150 m one-to-one to query centres within 2 m. Two models cover 84.0–85.5% of 9,684 boxes; seed 0 has 8,277 matches. The seed-0 median 3×3 count readings are 15.4,21.9,36.7,51.4,111.2 for GT counts 0,1–2,3–5,6–15,>15. Spearman correlation is 0.51 overall (0.42–0.47 by range), and AUC for zero-return versus returning boxes is 0.72 (0.67–0.72 by range); the models agree within 0.03. The same five-sweep field read at all GT centres yields 0.52/0.73 (correlation/AUC), versus 0.55/0.76 for a single sweep. On the matched subset, query-centre statistics are close to the GT-centre reference, providing the requested measurement of correspondence.

W2. Detection gains, auxiliary supervision and benchmark scope.

  • Detection: The original seven four-epoch baseline pairs remain +0.31 ± 0.36 NDS and −0.27 ± 0.22 mAP. The two completed eight-epoch pairs give +0.67/+1.19 NDS and +0.26/−0.13 mAP: +0.93 ± 0.26 NDS, +0.07 ± 0.20 mAP. The original held-out far-range recall result is unchanged (§4.4).
  • Density head (Q1): Three matched seeds give +0.67 ± 0.68 NDS (+1.26,+1.43,−0.69) and +0.11 ± 0.25 mAP. The auxiliary target encourages count information in the shared representation; the multi-seed ablation and loss definition are detailed in Part 2, Q1. Its output is unused for detection and the head can be omitted at inference.
  • Scope: The new controls use TruckScenes with the same detector and training recipe.
AuthorsOpenReview ↗

Response to Reviewer qhcn — Part 2/2: objective, branch isolation and equations

Q1. Density-loss weight, decoder aggregation, multi-seed results and Table 12.

  • Objective (Eq. 3, App. K): L=Ldet+0.1 BCE(zq,σ(4(nq−0.5)))\mathcal{L}=\mathcal{L}_{\rm det}+0.1\,\mathrm{BCE}(z_q,\sigma(4(n_q-0.5))), with nq=sg[E0(cq)]n_q=\mathrm{sg}[E_0(c_q)] (single-cell log-count at the decoded query centre). BCE is applied to logits and averaged over batch and queries. Stop-gradient blocks gradients through the target. The auxiliary loss is applied only at iteration 6, the final iteration of the weight-shared decoder; it is not summed six times.
  • Matched four-epoch ablation (Table 26, seeds 0–2): EviScope − no-density is +0.67 ± 0.68 NDS (+1.26,+1.43,−0.69) and +0.11 ± 0.25 mAP (+0.01,+0.59,−0.26), mean ± SE. The effect varies across seeds. Its output is unused for detection; the head can be omitted at inference.
  • Role: All three output heads have zero initial weights. After their first update, the density objective encourages the shared representation to retain count information; it does not supervise an optimal sampling policy. Without this branch, the module averages 26.14 mAP/38.72 NDS, or −0.03 ± 0.35 mAP and +0.48 ± 0.31 NDS versus baseline. Its +0.76 ± 0.19 mAP/+0.58 ± 0.14 NDS over geometry-only is descriptive: those variants differ in both evidence input and auxiliary loss.
  • Table 12: The one-epoch diagnostic enables multi-frame input from step 1; the four-epoch schedule enables it from epoch 2. Table 11's seed-0 density change (38.93→40.19 NDS, 26.46→26.47 mAP) and Table 12's opposite NDS change (27.58→22.67) therefore use different schedules and each one seed. Table 26 supplies the requested matched multi-seed comparison.

Q2. Fixed/adaptive radius with the same temporal branch; historical weights from current evidence.

  • Branch isolation: The fixed-0.6 control in §5.2 does not retain the same added temporal conditioner, so it is not the requested one-factor comparison. We report no completed fixed/adaptive-radius comparison with an otherwise identical temporal branch.
  • Conditioning rationale: The five-sweep field at the current keyframe measures local evidence availability, not which historical frame contains a hit. It conditions an allocation policy; the inherited cross-frame features provide alignment. The learned effective window exp⁡(H(πq))\exp(H(\pi_q)) lengthens as evidence falls in all 20/20 seed-by-range cells (Spearman −0.06 to −0.73), versus mixed signs in geometry-only (−0.26 to +0.37; Figure 7). This shows how current local evidence is associated with the learned temporal allocation.

Q3. Field, query extent, sampler integration and settings. These are explicit in §3.2, Algorithm 1 and Appendices K/U:

  • Field: E∈R5×128×128E\in\mathbb{R}^{5\times128\times128} uses five sweeps per radar over ±153.6\pm153.6 m (2.4 m cells). Channels 0–2 are log⁡(1+Cr)\log(1+C_r) for counts in (2r+1)×(2r+1)(2r+1)\times(2r+1) cells, r=0,1,2r=0,1,2; channels 3–4 are cell-mean RCS/32 and planar speed/10.
  • Descriptor (Eq. 1): ρq=∥cq,xy∥2\rho_q=\|c_{q,xy}\|_2, ρ~q=clip⁡(ρq/153.6,0,2)\tilde\rho_q=\operatorname{clip}(\rho_q/153.6,0,2) and sq=(wq+lq+hq)/3s_q=(w_q+l_q+h_q)/3 in metres. A bilinear read eq=E(cq)e_q=E(c_q) joins g(q)=[ρ~q,(ρq+1)−1,sq/(ρq+1),log⁡(1+sq)]g(q)=[\tilde\rho_q,(\rho_q+1)^{-1},s_q/(\rho_q+1),\log(1+s_q)] in a two-layer, 64-unit ReLU MLP yielding hqh_q.
  • Sampling (Eq. 2): αq=0.6+1.9σ(wα⊤hq+bα)\alpha_q=0.6+1.9\sigma(w_\alpha^\top h_q+b_\alpha) scales radial offsets: rq,k=dq+αqΔrℓ,q,kr_{q,k}=d_q+\alpha_q\Delta r_{\ell,q,k}. The image/BEV samplers' frame weights become (8πq,t)Aq,t(8\pi_{q,t})A_{q,t}, without renormalisation, with πq=softmax⁡(Wπhq+bπ)\pi_q=\operatorname{softmax}(W_\pi h_q+b_\pi). Zero output weights, zero frame biases and bα=logit⁡(0.4/1.9)b_\alpha=\operatorname{logit}(0.4/1.9) give αq=1\alpha_q=1 and 8πq,t=18\pi_{q,t}=1 initially.
  • Summary: Table 28 lists the comparison settings; Table 15 separates inherited and added components; Algorithm 1 shows one decoder iteration and the loss above specifies its final-iteration aggregation.

Thank you for the author’s response.

My main concern is whether the experimental data demonstrate the method's effectiveness. This concern has two sources: first, the paper uses a single benchmark and does not include more mainstream datasets such as nuScenes and VoD. Second, on this benchmark, the improvements in mAP and NDS are less than 1% and therefore very marginal.

I believe that the authors’ explanation and additional experimental data have alleviated my concern to some extent. In addition, I am very grateful for the author’s supplementary details, which help readers better understand and reproduce this work. Based on my own concerns alone, I am currently inclined to raise my score, but I also see that reviewer J8Lx and others have raised some very critical concerns. I look forward to seeing the other reviewers' responses and hope to discuss with them.

Reviewer_J8LxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1(−1)
Presentation
2(−1)
Soundness
1
Read review
Paper summary

EviScope adds a 5,450-parameter evidence-conditioned sampler to the RaCFormer radar–camera detector. A non-learned BEV radar-evidence field guides query-wise radial sampling and eight-frame temporal weights. Experiments on MAN TruckScenes examine multi-seed detection performance, evidence sensitivity, and long-range recall.

Critical strengths
  1. The paper is unusually transparent about negative results: seven matched-seed comparisons, an explicitly failed held-out subset test, and an evidence-blind control prevent overclaiming a marginal gain.
  2. The proposed conditioning module is lightweight, and the range-by-return analysis is a useful diagnostic of radar sparsity and detection difficulty.
Critical weaknesses
  1. The architectural novelty over RaCFormer is very limited. EviScope Figure 2 is strikingly similar to RaCFormer Figure 2, including the upstream modality branches, query layout, decoder structure, and visual organization. The manuscript itself describes the method as a small RaCFormer addition; the main difference is a 5.5K-parameter conditioner. The inherited architecture and genuinely new elements should be distinguished explicitly, and the incremental novelty better justified.
  2. The central module fails to demonstrate a reproducible detection gain. Across seven matched four-epoch pairs, EviScope changes NDS by only +0.31 ± 0.36 and mAP by −0.27 ± 0.22 (mean ± SE). The single eight-epoch pair (+0.67 NDS) cannot overturn this result; the stronger V2-99 backbone also yields −1.1 NDS in its one reported pair.
  3. The apparent advantage over published detectors largely comes from the module-free training recipe. The eight-epoch RaCFormer baseline already reaches 26.4 mAP / 39.2 NDS, compared with EviScope’s 26.7 / 39.9. Published comparisons are not compute-matched, so they do not establish an EviScope-specific advance.
  4. The core evidence-conditioning mechanism is not causally isolated. A full-budget, matched-seed geometry-only conditioner is missing. Replacing learned radial scales barely affects NDS (−0.03 ± 0.18), whereas uniform temporal weights cause a large drop even in the evidence-blind control (−11.75 NDS), consistent with co-adaptation rather than a demonstrated benefit of local radar evidence.
  5. The long-range motivation remains unverified. The claimed 100–150 m, low-return recall gain fails the authors’ own held-out-seed test (−0.67 and −0.59 points). Moreover, the motivating radar-return counts are measured within ground-truth boxes, while the model queries evidence at predicted centres; their correspondence is not established, and evaluation is limited to TruckScenes validation.
Questions for authors
  1. Can the authors directly compare EviScope Figure 2 with RaCFormer Figure 2, identify all inherited elements, and state precisely what is architecturally new beyond the small conditioner?
  2. Can the authors run full-schedule, multi-seed ablations with a properly trained geometry-only conditioner, temporal-only and radial-only branches, and shuffled radar-evidence inputs?
  3. Can the authors demonstrate a reproducible improvement over their own module-free baseline at matched training budgets, rather than relying on unmatched published results or a single eight-epoch run?
  4. Can the authors validate the link between ground-truth object-level radar counts and query-local evidence, and reproduce the targeted long-range improvements on held-out seeds or another dataset?
2 comments
AuthorsOpenReview ↗

Response to Reviewer J8Lx — Part 1/2: architecture and matched-budget results

Thank you for the direct review. We address W1/Q1, W2/Q3 and W3 here, and W4/Q2 and W5/Q4 in Part 2. All ± values below are mean ± SE over matched seeds.

W1 and Q1. Figure 2, inherited architecture and architectural novelty. EviScope is a 5,450-parameter conditioning module attached to RaCFormer. We revised Figure 2 and added Table 15 and Algorithm 1 to make this boundary explicit:

  • Figure 2: Panel (a) condenses the inherited detector and marks the attachment at ray sampling in the shared decoder. Panel (b) depicts the evidence field, query-centre read, MLP and output heads. The upstream branches and decoder layout in the original diagram were inherited, not contributions of EviScope.
  • Inherited components (Table 15): Camera encoder/FPN and LSS transform, radar pillar/BEV encoder, polar query formulation and denoising, six-iteration shared decoder, cross-modal ray sampling, adaptive mixing, prediction heads and detection losses.
  • Shared TruckScenes configuration: Four cameras, six radars, ±153.6 m extent, 256×256 detector BEV, 150×6=900 queries, class mapping, CBGS and eight-frame training (multi-frame input from epoch 2). Baseline and EviScope use the same configuration; these adaptations are not attributed to the new module.
  • Added components: A parameter-free 128×128×5 radar field; bilinear sampling at decoded query centres; a geometry/evidence descriptor and 64→64 MLP; a radial-offset multiplier in [0.6,2.5]; eight-frame weight multipliers; and an auxiliary density objective at the final decoder iteration. The output heads are initialised to identity radial/temporal modulation. The density prediction does not enter the detector output. The contribution is the explicit local evidence input and its controlled evaluation within this small conditioner, not an alternative backbone. The geometry-only control changes only access to that input.

W2 and Q3. Reproducible gain over the module-free baseline at matched budgets.

  • Four epochs (Table 7): The original seven pairs remain +0.31 ± 0.36 NDS and −0.27 ± 0.22 mAP (5/7 NDS positive).
  • Geometry-only comparison (Table 26): At seeds 0–2, its observed mean NDS is 38.14 versus baseline 38.24, with lower mAP in all three pairs (−0.49,−1.01,−0.86). Supplying evidence yields +0.87 ± 0.21 mAP (+1.12,+1.04,+0.46) and +1.25 ± 0.72 NDS (+1.62,+2.26,−0.14), with lower mATE in all three pairs. This supports the evidence input within the conditioner; it is a different comparison from module versus baseline.
  • Eight epochs (Table 26): Seed 1 gives 40.08 versus 38.89 NDS (+1.19) and 26.49 versus 26.62 mAP (−0.13), with both arms on the same node. Its mAOE/mAVE/mASE changes are 0.548→0.457, 0.743→0.719 and 0.420→0.407. Together with seed 0 (+0.67 NDS,+0.26 mAP), the two completed pairs average +0.93 ± 0.26 NDS and +0.07 ± 0.20 mAP. Both completed pairs show an NDS improvement over our own baseline at the same eight-epoch budget.
  • V2-99 (App. D): The single matched four-epoch pair remains −1.1 NDS and −0.1 mAP. The new matched controls concern ResNet-50.

W3. Published comparisons and the module-free recipe. The recipe is a separate contribution: the module-free ResNet-50 baseline averages 25.9 mAP (SD 0.45, seven runs) versus 24.1 for the best published entry. As you note, the eight-epoch baseline itself reaches 26.4/39.2 mAP/NDS, versus EviScope's 26.7/39.9 in that original pair. Published comparisons are not compute-matched and do not isolate the module. The added 256×864, eight-epoch EviScope row scores 28.2/39.9 versus the published ResNet-50 row's 23.6/37.4; it is explicitly an unpaired reference, not module-specific evidence (Table 1, App. C). Our module-effect conclusions use the matched comparisons above.

AuthorsOpenReview ↗

Response to Reviewer J8Lx — Part 2/2: causal controls and far-range evidence

W4 and Q2. Geometry-only, shuffled-evidence, temporal-only and radial-only controls. The completed full-schedule evidence is summarised below (four epochs; mean ± SE over matched seeds; Appendix T, Table 26).

ControlCompleted matched seedsEviScope − control, mAP / NDS
Geometry-only0–2+0.87 ± 0.21 / +1.25 ± 0.72
Shuffled evidence0–1+0.74 ± 0.54 / +2.44 ± 0.45
  • Geometry-only: It retains the same 5,450 parameters, query geometry, learned radial/temporal heads, identity initialisation and auxiliary target, zeroing only the five evidence inputs. Differences are +1.12,+1.04,+0.46 mAP and +1.62,+2.26,−0.14 NDS. At seed 2, mATE improves 0.662→0.648, while orientation/velocity errors offset that improvement in NDS. Same-node re-evaluation gives +0.91 ± 0.22 mAP and +1.26 ± 0.73 NDS (Table 26†).
  • Shuffled evidence: Randomly permuting vectors across each sample's 900 queries in training and evaluation preserves the frame-level distribution while disrupting query alignment. Seed 0 scores 25.19/37.30 and seed 1 scores 26.31/38.25 (mAP/NDS). EviScope exceeds them by +1.28/+2.89 and +0.20/+1.99. The comparison supports the value of query-aligned evidence within the same conditioner.
  • Branch isolation: The completed controls above isolate the evidence input. We do not report completed full-schedule temporal-only/radial-only comparisons. The fixed-0.6 control lacks the same added temporal conditioner; inference-time substitutions address co-adaptation (§5.2).
  • Allocation (Figure 7): Within every range band, radius decreases with evidence (Spearman −0.89 to −0.95, five EviScope seeds), and temporal windows shorten in all 20 seed-by-band cells. Geometry-only radii stay at 0.600–0.607 (SD <0.001). This allocation pattern repeats across seeds and range bands.

W5 and Q4. Object-count/query-evidence correspondence and far-range replication.

  • Correspondence (App. S, Figure 5, Table 25): Across 300 validation frames and two models, 84.0–85.5% of 9,684 GT boxes have a one-to-one query match within 2 m (8,277 matches for seed 0). Median query-centre 3×3 count readings rise through 15.4,21.9,36.7,51.4,111.2 for GT counts 0,1–2,3–5,6–15,>15. Correlation is 0.51 overall (0.42–0.47 by range); zero-return discrimination has AUC 0.72, including at 100–150 m. The five-sweep field read at all GT centres gives correlation/AUC 0.52/0.73, versus 0.55/0.76 for one sweep. On the matched subset, the query result is close to this GT-centre reference, providing the requested correspondence measurement.
  • Far-range replication: The original held-out recall result remains −0.67 and −0.59 points (§4.4). The new probe reproduces the evidence/radius association at 100–150 m in all five EviScope seeds (Spearman −0.89 to −0.91; App. T); this is an allocation result, separate from held-out detection improvement. The completed evaluations use TruckScenes.
Reviewer_mFjWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EviScope, an evidence-conditioned sampling module for long-range camera-radar 3D detection. It builds a BEV evidence field from current-frame radar returns and uses query-local evidence to predict radial sampling scales and temporal fusion weights, with auxiliary density supervision. The proposed EviScope achieves promising performance on the TruckScenes dataset

Critical strengths
  1. This paper proposes EviScope, which combines a parameter-free radar evidence field with query geometry to jointly modulate the radial sampling extent and temporal aggregation weights.

2.The proposed EviScope achieves promising performance on the TruckScenes dataset.

Critical weaknesses

1.The novelty of EviScope is limited. Compared with RaCFormer, it incorporates additional contextual information to guide sampling and potentially improve its performance, but this comes at the cost of additional parameters and computation.

2.Can the proposed method be evaluated on the VoD dataset to further assess its effectiveness and generalizability?

  1. It is encouraged that the authors compare their method with recent approaches, including HGFusion and RCBEVDet++.
Questions for authors

1.It is recommended that the authors further clarify the differences between the proposed EviScope and RaCFormer to better highlight the novelty of this work.

2.Can the proposed method be evaluated on the VoD dataset to further assess its effectiveness and generalizability?

  1. It is encouraged that the authors compare their method with recent approaches, including HGFusion and RCBEVDet++.
1 comment
AuthorsOpenReview ↗

Response to Reviewer mFjW: novelty, cost, scope and related methods

Thank you for the review. We address the corresponding weaknesses and questions together below. All ± values are mean ± SE over matched seeds.

W1 and Q1. Novelty relative to RaCFormer and added cost.

  • Architecture (Figure 2, Eqs. 1–3, Table 15): RaCFormer supplies the modality encoders, polar queries, shared decoder, ray sampling and prediction heads. EviScope adds a 5,450-parameter conditioner driven by an explicit non-learned radar summary. Five sweeps per radar form 1×1/3×3/5×5 log-counts, mean RCS and planar speed on a BEV grid. Bilinear reads at decoded query centres modulate radial offsets through αq∈[0.6,2.5]\alpha_q\in[0.6,2.5] and eight-frame weights through 8πq8\pi_q. Figure 2 separates this addition from the inherited detector; Table 15 separates both from shared TruckScenes adaptations.
  • Geometry-only control (Table 26, four epochs, seeds 0–2): Zeroing evidence inputs leaves the same parameters, heads and auxiliary target. Its observed mean NDS is 38.14 versus baseline 38.24; mAP is lower than baseline at each seed. EviScope − geometry-only is +0.87 ± 0.21 mAP (+1.12,+1.04,+0.46) and +1.25 ± 0.72 NDS (+1.62,+2.26,−0.14), with lower translation error at every seed. Thus the comparison tests the evidence input within an otherwise fixed conditioner.
  • Shuffled evidence (seeds 0–1): Permuting evidence across queries in training/evaluation gives controls at 25.19/37.30 and 26.31/38.25 mAP/NDS. EviScope exceeds them by +1.28/+2.89 and +0.20/+1.99, averaging +0.74 ± 0.54 mAP and +2.44 ± 0.45 NDS. The matched test supports the contribution of query-aligned evidence.
  • Baseline comparison: Seven four-epoch pairs remain +0.31 ± 0.36 NDS and −0.27 ± 0.22 mAP. Two completed eight-epoch pairs give +0.67/+1.19 NDS, averaging +0.93 ± 0.26 NDS and +0.07 ± 0.20 mAP.
  • Cost (App. J): The module adds 5,450 parameters (0.009% of 64.0M) and analytically 5,312 MACs/query/iteration (about 29M/sample), plus construction of the non-learned grid. In our one-A100, batch-one eight-frame measurement after warm-up, median wall-clock latency over 50 iterations was 965 ms versus 997 ms for baseline; both reported 3.811 GB peak memory.

W2 and Q2. VoD evaluation and generalisability. We focus on TruckScenes for the 0–150 m regime: 72.6% of cars at 100–150 m have zero radar returns (§3.1), whereas VoD's official annotated range is 0–50 m in the camera FoV. RaCFormer's VoD configuration is compatible with a shorter-range evaluation, but we report no completed VoD experiment. Our additions answer the evidence-input question with matched TruckScenes controls. Across the official TruckScenes scene tags, mean ΔNDS over geometry-only is positive in 19/20 subsets; all three seeds are positive in 7/20 (Table 27). These are overlapping within-dataset conditions.

W3 and Q3. Recent methods: HGFusion and RCBEVDet++. We interpret HGFusion as HGSFusion (Gu et al., AAAI 2025). Section 2 now discusses it and RCBEVDet++ (Lin et al., arXiv:2409.04979). The comparison is architectural; it is not a new matched benchmark ranking.

MethodReported benchmarksDistinction
RaCFormer (CVPR 2025)nuScenes, VoDPolar/sector ray sampling and temporal mixing from query features; our base detector
HGSFusion (AAAI 2025)VoD, TJ4DRadSetHybrid radar point generation and radar–camera feature alignment
RCBEVDet / RCBEVDet++nuScenes, VoD (RCBEVDet); nuScenes (++)Radar encoding and cross-modal fusion; ++ extends to query-based sparse fusion
EviScopeTruckScenesKeeps radar returns and the detector, adding explicit local radar statistics to modulate sampling and frame weights

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: completed controls and manuscript revisions

We thank all four reviewers. The central question is whether explicit query-local radar evidence helps beyond query geometry and added adaptive capacity. The revised manuscript answers this question with full-schedule matched controls, a query-evidence correspondence measurement, and replicated allocation probes. All ± values below are mean ± SE across matched seeds.

  1. Geometry-only control, four epochs, seeds 0–2 (Table 26). Zeroing the five evidence inputs preserves the 5,450 parameters, learned radial/temporal heads, identity initialisation, and auxiliary target. Its observed mean NDS is 38.14 versus the module-free baseline's 38.24, with lower mAP in all three pairs (−0.49, −1.01, −0.86). EviScope − geometry-only is +0.87 ± 0.21 mAP (+1.12, +1.04, +0.46) and +1.25 ± 0.72 NDS (+1.62, +2.26, −0.14), with lower translation error in all three pairs. Same-node re-evaluation gives +0.91 ± 0.22 mAP and +1.26 ± 0.73 NDS (Table 26†).
  2. Shuffled-evidence control, four epochs, seeds 0–1 (Table 26). Permuting each sample's evidence across its 900 queries in training and evaluation preserves the frame-level distribution but breaks query alignment. The controls score 25.19/37.30 and 26.31/38.25 (mAP/NDS). EviScope exceeds them by +1.28/+2.89 and +0.20/+1.99, averaging +0.74 ± 0.54 mAP and +2.44 ± 0.45 NDS. The matched comparison supports the value of query-aligned radar evidence.
  3. Eight-epoch baseline pairs, seeds 0–1 (Table 26). Both completed pairs are positive on NDS: +0.67 and +1.19, averaging +0.93 ± 0.26 NDS, with +0.07 ± 0.20 mAP. This adds a second positive NDS pair to the original eight-epoch result. The original seven four-epoch comparisons are reported separately in Table 7.

Guide to manuscript revisions:

LocationAdded or clarified content
Abstract, §§1–2, §6Claims, related work, and conclusions updated to the controls
§3.2, Eqs. 1–3, Fig. 2Evidence field, geometry, sampling and loss equations; redrawn architecture
Table 1, Apps. C and J–K; Table 15, Algorithm 1Resolution reference, measured overhead, inherited/added components and pseudocode
App. S, Fig. 5, Table 25Object-count/query-evidence correspondence
App. T, Table 26; Table 11Geometry-only, shuffled-evidence, density-head and eight-epoch comparisons
App. T, Figs. 6–7, Table 27Paired summary, allocation probe and scene-condition comparisons
App. U, Table 28Summary of experimental settings

Substantive changes are blue. Table 26 distinguishes results already present in the original submission (black, with source-table references) from newly reported results (blue); newly reported does not necessarily mean newly trained.

What the diagnostics establish. Within each range band, EviScope's radius decreases with local evidence (Spearman −0.89 to −0.95 across five seeds), and its temporal window shortens in all 20 seed-by-band cells; geometry-only radii stay at 0.600–0.607. Matched queries achieve Spearman 0.51 and AUC 0.72 for object-level evidence, with coverage 84.0–85.5% (App. S). Together with the trained controls, these measurements connect the added input to the learned allocation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.