Research map / Point clouds / Segmentation and detection
Not All Points Are Equal: Learning Highly Efficient Point-Based Detectors for 3D LiDAR Point Clouds
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace D-FPS in the later sampling layers with instance-aware sampling (ablation) | IA-SSD · KITTI validation, Pedestrian and Cyclist | 3D mAP (11 recall positions): 50.46 -> 58.81; 65.19 -> 70.15 | Helped |
| Combine class-aware and centroid-aware sampling (ablation) | IA-SSD · KITTI validation, Cyclist | 3D mAP (11 recall positions): 70.15 -> 71.33 | Helped |
| Use centroid-aware sampling only rather than mixed class- and centroid-aware sampling (ablation) | IA-SSD · KITTI validation, Pedestrian and Car | 3D mAP (11 recall positions): 58.49 -> 58.91; 79.54 -> 79.57 | Helped |
| Use centroid-aware sampling only rather than mixed sampling (ablation) | IA-SSD · KITTI validation, Cyclist | 3D mAP (11 recall positions): 71.33 -> 71.24 | Hurt |
| Include contextual points in centroid prediction instead of using centers-assign (ablation) | IA-SSD · KITTI validation, Pedestrian and Cyclist | 3D mAP (11 recall positions): 24.90 -> 58.91; 37.12 -> 71.24 | Helped |
| Expand centroid context by a 2× box factor rather than +1.0m (ablation) | IA-SSD · KITTI validation, Cyclist | 3D mAP (11 recall positions): 71.24 -> 68.16 | Hurt |
| Use instance-aware sampling in early encoding layers (ablation) | IA-SSD · KITTI validation | 3D detection performance | Hurt |
| Preserve more foreground points through instance-aware sampling (ablation) | IA-SSD · KITTI validation, especially small objects | 3D detection performance | Helped |
| Train one multi-class model rather than separate class-specific models | IA-SSD · KITTI test | 3D AP, Cyclist Moderate: 66.01 -> 61.94 | Hurt |
| Use IA-SSD rather than PV-RCNN | 3D detection on Waymo validation · Pedestrian Level 1 | mAP: 62.66 -> 69.38 | Helped |
| Use IA-SSD rather than PV-RCNN | 3D detection on Waymo validation · Vehicle Level 1 | mAP: 74.06 -> 70.53 | Hurt |
| Use IA-SSD rather than Feat-FPS for sampling | Point sampling at 16384 points · Sampling efficiency | Time: 40.6 ms -> 0.5 ms | Helped |
Closest papers by meaning
- PillarNeXt: Rethinking Network Designs for 3D Object Detection in LiDAR Point Clouds (CVPR 2023 · similarity 0.91)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.91)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.90)
- AS-Det: Active Sampling for Adaptive 3D Object Detection in Point Clouds (AAAI 2025 · similarity 0.89)
- Beyond Virtual Points: Depth-Enhanced LiDAR-only 3D Object Detection with Semi-Supervised Learning (Student Abstract) (AAAI 2025 · similarity 0.89)
- LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection (ICLR 2024 · similarity 0.89)
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.