Research map / Detection and segmentation / 3D detection and driving
Rethinking Pseudo-LiDAR Representation
Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, Wanli Ouyang
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 |
|---|---|---|---|
| represent generated 3D data as image patches processed by 2D CNN instead of point cloud with PointNet | pseudo-LiDAR · KITTI val Car monocular | AP_3D R11 Moderate: 18.4 -> 18.4 | No clear effect |
| use world coordinates (x,y) instead of image coordinates (u,v) (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 15.7 -> 22.0 | Improved |
| add the y coordinate to the {x,z} input (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 18.3 -> 22.0 | Improved |
| add x and y coordinates to depth-only input (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 3.48 -> 22.0 | Improved |
| use 2D CNN backbone ResNet-18 instead of PointNet-18 (ablation) | PatchNet-vanilla · KITTI val Car monocular | AP_3D R11 Hard: 17.0 -> 19.1 | Improved |
| use SE-ResNet-18 instead of ResNet-18 (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 21.3 -> 21.5 | No clear effect |
| use ResNeXt-101 instead of ResNeXt-50 (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Hard: 17.3 -> 17.0 | Hurt |
| add mask global pooling (max) instead of standard global max pooling (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Easy: 32.4 -> 35.1 | Improved |
| use mask max pooling instead of mask average pooling (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 21.6 -> 22.0 | Improved |
| add instance assignment with difficulty-specific heads (distance switcher) (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Easy: 33.7 -> 35.1 | Improved |
| use distance instead of predicted difficulty as switcher (ablation) | PatchNet · KITTI val Car monocular | AP_3D R11 Moderate: 22.1 -> 22.0 | No clear effect |
| use PatchNet instead of AM3D on test set Easy | monocular 3D detection · KITTI test Car | AP_3D R40 Easy: 16.50 -> 15.68 | Hurt |
Closest papers by meaning
- End-to-End Pseudo-LiDAR for Image-Based 3D Object Detection (CVPR 2020 · similarity 0.91)
- Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving (CVPR 2019 · similarity 0.90)
- Pseudo-Stereo for Monocular 3D Object Detection in Autonomous Driving (CVPR 2022 · similarity 0.89)
- LIGA-Stereo: Learning LiDAR Geometry Aware Representations for Stereo-Based 3D Detector (ICCV 2021 · similarity 0.88)
- LiDAR R-CNN: An Efficient and Universal 3D Object Detector (CVPR 2021 · similarity 0.88)
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (AAAI 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.
