Research map / Point clouds / Segmentation and detection
GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure
Bin Yang, Mohamed Abdelsamad, Miao Zhang, Michael Ulrich, Yakov Miron, Abhinav Valada, Alexandru Paul Condurache
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 hallucinated ghost-point targets to the predictor | PointINS · nuScenes (decoder probing) | mAP: 56.7 -> 59.5 | Improved |
| add hallucinated ghost-point targets to the predictor | PointINS · nuScenes (decoder probing) | NDS: 62.5 -> 64.2 | Improved |
| add hallucination pretraining then fine-tune | PTv3 · nuScenes (full fine-tuning) | mAP: 63.8 -> 67.5 | Improved |
| restrict predictor to masked observed points (no hallucination) (ablation) | PointINS · nuScenes (decoder probing) | mAP: 56.7 -> 57.3 | Improved |
| replace ghost targets with zero offsets and uniform Softmaps (ablation) | PointINS · nuScenes (decoder probing) | mAP: 57.3 -> 57.6 | No clear effect |
| add informative hallucinated targets rather than zero targets (ablation) | GhostPoint · nuScenes (decoder probing) | mAP: 57.6 -> 59.5 | Improved |
| swap masked points for the same number of hallucinated points (ablation) | GhostPoint · nuScenes (decoder probing) | mAP: 59.4 -> 59.5 | No clear effect |
| add neighborhood sampling with KNN interpolation, then predictor with warmup (ablation) | GhostPoint · nuScenes (decoder probing) | mAP | Improved |
| add an auxiliary box regression target (ablation) | GhostPoint · nuScenes (decoder probing) | mAP | Hurt |
| add KNN-based token initialization for teacher ghost tokens (ablation) | GhostPoint · nuScenes (decoder probing) | mAP | Improved |
| increase predictor depth beyond 3 blocks (ablation) | GhostPoint · nuScenes (decoder probing) | mAP | No clear effect |
| add GhostPoint pretraining under 0.1% labels | PTv3 · Waymo (0.1% labels) | mAP: 20.0 -> 28.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SONATA — Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Unsupervised Object Detection From LiDAR Point Clouds (CVPR 2023 · similarity 0.92)
- Point Density-Aware Voxels for LiDAR 3D Object Detection (CVPR 2022 · similarity 0.91)
- Image-to-Lidar Self-Supervised Distillation for Autonomous Driving Data (CVPR 2022 · similarity 0.91)
- ALSO: Automotive Lidar Self-Supervision by Occupancy Estimation (CVPR 2023 · similarity 0.90)
- Pre-training LiDAR-based 3D Object Detectors through Colorization (ICLR 2024 · similarity 0.89)
- Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · 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.
