Research map / Point clouds / Segmentation and detection
Clusterformer: Cluster-based Transformer for 3D Object Detection in Point Clouds
Yu Pei, Xian Zhao, Hao Li, Jingyuan Ma, Jingwei Zhang, Shiliang Pu
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 |
|---|---|---|---|
| replace cluster feature pooling with intra-cluster cross-attention (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 68.6 -> 70.0 | Improved |
| add cluster-wise self-attention between queries (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 70.0 -> 70.4 | Improved |
| add Query2Key strategy (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 70.4 -> 71.2 | Improved |
| initialize queries from cluster centers instead of farthest point sampling (ablation) | Clusterformer · Waymo val (ablation), vehicle | L2 mAPH: 61.7 -> 65.3 | Improved |
| limit interaction to the same cluster rather than distance-wise range (ablation) | Clusterformer · Waymo val (ablation), vehicle | L2 mAPH: 65.3 -> 66.9 | Improved |
| global-wise interaction range for queries (ablation) | Clusterformer · Waymo val (ablation), vehicle | L2 mAPH: 66.9 -> 49.4 (cluster center init) | Hurt |
| zero-initialized queries with intra-cluster interaction (ablation) | Clusterformer · Waymo val (ablation), vehicle | L2 mAPH: 66.9 -> 65.9 | Mixed |
| increase decoder layers with Query2Key (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: +1.6 with Query2Key vs +0.8 without (2 to 4 layers) | Improved |
| remove the local maximum operation in cluster generation (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 71.2 -> 69.9 | Hurt |
| remove NMS with Max-IoU matching (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 71.2 -> 68.1 | Hurt |
| use Hungarian matching so NMS is not needed (ablation) | Clusterformer · Waymo val (ablation) | Overall L2 mAPH: 68.1 -> 70.2 (no NMS) | Improved |
| Clusterformer versus Centerformer in latency and memory | Clusterformer · Waymo val | latency / inference memory: 123ms -> 103ms; 10.4G -> 5.7G | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Centerformer — CenterFormer: Center-based Transformer for 3D Object Detection (ECCV 2022) · better on 3 of 3 shared results
- PV-RCNN++ — PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection (CVPR 2020) · 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
- 3D Object Detection With Pointformer (CVPR 2021 · similarity 0.90)
- Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection From Point Clouds (CVPR 2022 · similarity 0.89)
- An End-to-End Transformer Model for 3D Object Detection (ICCV 2021 · similarity 0.88)
- Voxel Transformer for 3D Object Detection (ICCV 2021 · similarity 0.88)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.86)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · 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.
