Research map / Detection and segmentation / 3D detection and driving
StereoDistill: Pick the Cream from LiDAR for Distilling Stereo-Based 3D Object Detection
Zhe Liu, Xiaoqing Ye, Xiao Tan, Errui Ding, Xiang Bai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Add response-level XGD and CLD to the stereo detector (ablation) | DSGN with Improved LIGA features · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: +2.72 / +6.65 / +1.41 | Improved |
| Add XGD without CLD (ablation) | Improved LIGA · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 86.62 -> 86.78 (Car Easy); 40.11 -> 45.37 (Pedestrian Moderate) | Improved |
| Add CLD without XGD (ablation) | Improved LIGA · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 67.03 -> 67.57; 40.11 -> 40.69; 40.90 -> 41.65 | Improved |
| Use center, size, and angle components together rather than one component alone (ablation) | XGD · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 67.84 / 44.11 / 41.18 -> 69.75 / 46.76 / 42.31 | Mixed |
| Use center component only for XGD (ablation) | XGD · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 68.22 / 46.33 / 42.19 | Mixed |
| Use all positive anchors rather than confidence-filtered high-quality boxes (ablation) | XGD · KITTI validation, moderate Cars and Pedestrians | 3D AP: 67.72 / 43.65 -> 69.75 / 46.76 | Improved |
| Distill all foreground positions rather than only positive-anchor positions (ablation) | CLD · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 67.66 / 45.71 / 42.32 -> 69.75 / 46.76 / 42.31 | Mixed |
| Aggregate anchors at each position into a unified distribution instead of treating anchors separately (ablation) | CLD · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: 67.74 / 46.18 / 40.69 -> 69.75 / 46.76 / 42.31 | Improved |
| Replace LIGA with StereoDistill response-level distillation | DSGN · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D mAP: 48.20 -> 50.12 | Improved |
| Add feature-based distillation and response-based distillation together (ablation) | DSGN · KITTI validation, moderate Car/Pedestrian/Cyclist | Mean 3D AP: 42.60 -> 52.94 | Improved |
| Distill with XGD and CLD using PointPillars as the LiDAR teacher | LIGA and Improved LIGA · KITTI validation, moderate Car/Pedestrian/Cyclist | 3D AP: +1.28 / +2.09 / +1.84; +1.30 / +3.35 / +1.72 | Improved |
| Use StereoDistill instead of LIGA | DSGN · Argoverse validation, Car and Pedestrian | 3D AP: 34.37 -> 37.55; 10.76 -> 13.70 | Improved |
Closest papers by meaning
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.91)
- Advancing 3D Object Detection With Depth-Aware Spatial Knowledge Distillation (TPAMI 2025 · similarity 0.91)
- X3KD: Knowledge Distillation Across Modalities, Tasks and Stages for Multi-Camera 3D Object Detection (CVPR 2023 · similarity 0.90)
- RayD3D: Distilling Depth Knowledge Along the Ray for Robust Multi-View 3D Object Detection (AAAI 2026 · similarity 0.90)
- LabelDistill: Label-guided Cross-modal Knowledge Distillation for Camera-based 3D Object Detection (ECCV 2024 · similarity 0.89)
- Weak-to-Strong 3D Object Detection with X-Ray Distillation (CVPR 2024 · similarity 0.88)
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.
