SGDet3D++: Geometry-Grounded Semantics for 4D Radar and Camera 3D Object Detection

1 INTRODUCTION

• Projection and proximity alone cannot establish whether image or radar evidence supports a query, since tokens may describe occluders and returns may come from multipath or reflectors outside the object.The introduction frames evidence–hypothesis compatibility as a decision distinct from scene-level alignment.Source: §1 · p. 1
• SGDet3D++ conditions or verifies radar evidence using each sparse query’s evolving 3D anchor before semantic retrieval, geometric refinement, and temporal inheritance.This extends SGDet3D’s scene-level BEV fusion and projected supports toward explicit object-level decisions.Source: §1 · p. 2
• The proposed decoder contract separates candidate access from evidence admission: the evolving 3D state establishes semantic, geometric, or motion compatibility before evidence changes the query.Source: §1 · p. 2
• AGR conditions image retrieval with pooled anchor-consistent radar statistics, GCR reads individual object-owned returns for box refinement, and DVC admits pose-aligned history with current radial motion.Source: §1 · p. 2
• The paper reports validation of accuracy and operating mechanisms on three benchmarks, with controlled ablations and held-out diagnostics of retrieval redistribution, return ownership, and memory inheritance.Source: §1 · p. 2

2.1 4D RADAR-BASED 3D DETECTION

• 4D radar provides sparse returns with range, reflectivity, and motion, but limited class semantics and incomplete object boundaries motivate using it as object-level evidence rather than a dense semantic substitute for images.Source: §2.1 · p. 2

2.2 4D RADAR AND CAMERA FUSION 3D DETECTION

• SGDet3D++ extends scene-level fusion by associating each query with anchor-consistent radar geometry before semantic retrieval, return-level refinement, and Doppler-verified temporal inheritance.The cited prior approaches include radar–image proposal association, object-query sampling, and proposal refinement.Source: §2.2 · p. 3

3.1 OVERVIEW

• The decoder updates each object hypothesis from image and radar features, raw returns, and temporal memory, with the current anchor defining evidence support before it modifies the state.Rather than adding another fused scene representation, it selects evidence separately for each object.Source: §3.1 · p. 3

3.2 HYPOTHESIS-CONDITIONED EVIDENCE GROUNDING

• The yaw-aligned box and radius gates select anchor-consistent radar support, excluding returns outside the anchor or beyond its association range.The radius gate also prevents large or uncertain anchors from collecting distant returns.Source: §3.2 · p. 5
• AGR pools associated radar returns to condition image retrieval, while leaving the box-conditioned image support locations fixed.Its pooled descriptor includes local-coordinate means and return count; reported experiments use a 768-return budget and 256-dimensional decoder features.Source: §3.2 · p. 5
• GCR preserves associated returns as individual tokens, letting the current object feature attend to evidence for box refinement.Return-level geometry, strength, quality, age, and Doppler inform attention; the resulting context enters the ordinary refiner, while absent support leaves the original refinement path unchanged.Source: §3.2 · pp. 5–6
• DVC compares a nearby return’s ego-motion-compensated radial velocity with the pose-aligned anchor’s predicted radial velocity, replacing history only when mismatch exceeds a threshold.Missing radar evidence preserves history; contradicted slots are replaced by freshly initialized current features and anchors, without learnable parameters or a second cache.Source: §3.2 · p. 7

3.3 DETECTION OBJECTIVE

• The detection objective combines weighted detection supervision across decoder refinement layers with a multi-scale corrected-depth objective.The detection head supervises classification, box regression, centre quality, yaw quality, and planar velocity.Source: §3.3 · p. 7

4 EXPERIMENTS

• Evaluation covers OmniHD-Scenes and ManTruckScenes, using their respective detection metrics and shared validation splits, synchronized inputs, spatial range, and official evaluator.ManTruckScenes includes commercial-vehicle scenes; the controlled Base result in Table 5 is the mean of three independent runs.Source: §4 · pp. 7–8

4.2 ROBUSTNESS AND DECISION-LEVEL EVIDENCE

• At ±10° and ±1.0 m, SGDet3D++ retains 41.70 mAP and 52.76 ODS, exceeding Control by 10.67 and 12.56 points.On the adverse OmniHD-Scenes subset, SGDet3D++ obtains the best mAP and ODS.Source: Table 3 · p. 8, §4.2 · p. 8
• Target-return purity rises from 29.95% to 58.87% and foreign returns fall from 7.15% to 1.16%; with conflicts at 0.31% of eligible histories, DVC reaches 10.61% F1.DVC preserves 96.11% of consistent histories and retains 75.90% conflict recall. AGR shifts 8.68% of retrieval probability and changes the maximum-weight camera–level–support route for 15.77% of feature groups.Source: §4.2 · p. 9

4.3 ABLATION STUDY

• AGR and GCR together add 5.25 mAP, and DVC adds another 3.19 mAP; every component improves mAP and ODS over Base.The default configuration reaches 48.87 mAP, 58.22 ODS, 42.28 strict AP@1m, and 0.3350 mAVE; removing the yaw-aligned box test loses 1.89 mAP and 2.08 strict AP@1m. Joint local XYZ, RCS, and Doppler yields the best evaluated cue combination at 48.87 mAP and 58.22 ODS.Source: Table 5 · p. 9, §4.3 · p. 9
• SGDet3D++ attains the best mAP and NDS with 66.2 ms latency, 4.2 GB memory, 60.7 M parameters, and 306.3 GFLOPs.It also has the lowest latency and GFLOPs among the compared methods.Source: Table 7 · p. 9, §4.3 · p. 9

• SGDet3D++ grounds each query’s semantic retrieval, geometric refinement, and temporal inheritance in object-conditioned 4D radar evidence.Across two benchmarks, it improves detection, localization, and motion over the strongest baselines; robustness and resource analyses further show reliable and efficient detection.Source: §5 · p. 9

• Generative AI tools were used for literature organization, reviewer-style feedback, and language editing; the authors verified the work’s citations, technical descriptions, analyses, and revisions.The authors take responsibility for the final content.Source: AI USE STATEMENT · p. 10