Research map / Point clouds / Segmentation and detection
AS-Det: Active Sampling for Adaptive 3D Object Detection in Point Clouds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AS-Det, a point-based single-stage 3D detector with a learnable active sampling strategy (weighted random sampling with straight-through gradients and a Focal-KL sampling loss) plus multi-scale center feature aggregation, evaluated on KITTI, nuScenes, VoD and a self-collected 4D Radar dataset.
Problem
Point-based 3D object detectors rely on simplistic downsampling (FPS, feature-based, segmentation-based) that does not adapt across diverse point cloud types such as LiDAR and 4D Radar, which differ in density, FOV, and attributes.
Approach
Treat point sampling as a learnable decision: predict a per-point value distribution, sample via weighted random sampling, supervise it with an instance-aware, density-balanced objective using a Focal-KL divergence, and backpropagate detection loss through the sampler using the straight-through estimator.
Key results
- On KITTI test Car 3D moderate, AS-Det reaches 82.41% AP, +0.88% over FARP-Net and +2.09% over IA-SSD.
- On VoD (4D Radar) val, AS-Det reaches 77.73 mean AP in the Driving Corridor setting, +8.01% over SMURF, and 70.33 to 76.97 mAP gains come from active sampling over D-FPS in the encoder ablation.
- On nuScenes val, AS-Det achieves 51.92 mAP and 61.80 NDS, outperforming IA-SSD* (37.01 mAP) by more than 10 mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AS-Det instead of FARP-Net | Point-based 3D detection · KITTI test, Car 3D | APR40 mean: 82.96 -> 83.35 | Helped |
| use AS-Det instead of PVT-SSD | Car BEV detection · KITTI test | APR40 mean: 91.10 -> 90.93 | Hurt |
| use AS-Det instead of IA-SSD* | Point-based detection · nuScenes val | mAP: 37.01 -> 51.92 | Helped |
| use AS-Det instead of SMURF | 4D Radar detection · VoD val, Entire Annotated Area | APR40 mean: 50.97 -> 56.50 | Helped |
| replace D-FPS with active sampling in layers 3–4 (ablation) | AS-Det · KITTI val, three-class | APR40 mean: 69.50 -> 73.83 | Helped |
| replace D-FPS with active sampling in layers 2–4 (ablation) | AS-Det · KITTI val, three-class | APR40 mean: 69.50 -> 75.25 | Helped |
| replace D-FPS with active sampling in layers 3–4 (ablation) | AS-Det · VoD val, three-class | APR40 mean: 70.33 -> 76.97 | Helped |
| replace KL with FKL divergence for sampling loss (ablation) | AS-Det · KITTI val | mAP: 74.10 -> 75.25 (+1.15) | Helped |
| replace KL with FKL divergence for sampling loss (ablation) | AS-Det · nuScenes val | mAP: 48.03 -> 51.92 (+3.89) | Helped |
| replace single-scale VoteHead with two-scale center feature aggregation (ablation) | AS-Det · KITTI val | APR40: 72.26 -> 73.06 | Helped |
| increase center feature aggregation from two scales to three (ablation) | AS-Det · KITTI val | APR40: 73.06 -> 75.25 | Helped |
Surprising findings
- Seg-based sampling degrades significantly when more layers replace D-FPS, because density-imbalanced early layers let easy objects dominate the sampled subset.
- On the self-collected 4D Radar data, the KL-alike and weighted random sampling tolerate radar artifacts better than BCE-supervised segmentation-based sampling.
Limitations
- Performance on large-scale point clouds still lags voxel-based methods, per the authors.
- Self-collected 4D Radar dataset is not stated to be released.
- Sampling-based approach relies on a sampling objective built from ground-truth boxes.
Experimental setup
Datasets: KITTI nuScenes VoD Self-collected 4D Radar dataset
Benchmarks: KITTI nuScenes VoD Self-collected 4D Radar dataset
Metrics: AP R40 mAP NDS runtime (ms)
Baselines: SECOND SA-SSD PVGNet RDIoU VoxSet PVT-SSD PointRCNN Point-GNN 3DSSD IA-SSD FARP-Net PointPillars
Builds on: 3DSSD IA-SSD VoteNet PointNet++ Straight-Through estimator
Closest papers by meaning
- Exploring Active 3D Object Detection from a Generalization Perspective (ICLR 2023 · similarity 0.91)
- STONE: A Submodular Optimization Framework for Active 3D Object Detection (NeurIPS 2024 · similarity 0.88)
- V-DETR: DETR with Vertex Relative Position Encoding for 3D Object Detection (ICLR 2024 · similarity 0.88)
- Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 2024 · similarity 0.87)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.87)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.