Research map / Detection and segmentation / 3D detection and driving
FocalFormer3D: Focusing on Hard Instance for 3D Object Detection
Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, Jose M. Alvarez
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 one extra hard-instance-probing stage at 200 queries (ablation) | FocalFormer3D · nuScenes val | mAP: 65.3 -> 66.0 | Improved |
| increase single-stage queries from 200 to 600 without HIP (ablation) | FocalFormer3D · nuScenes val | mAP: 65.3 -> 65.4 | No clear effect |
| use 3 HIP stages with 600 queries instead of 1 stage (ablation) | FocalFormer3D · nuScenes val | mAP: 65.4 -> 66.5 | Improved |
| add pooling-based positive masking (ablation) | FocalFormer3D · nuScenes val | mAP: 65.3 -> 66.5 | Improved |
| use point masking instead of no masking (ablation) | FocalFormer3D · nuScenes val | mAP: 65.3 -> 65.9 | Improved |
| use box masking instead of pooling-based masking (ablation) | FocalFormer3D · nuScenes val | mAP: 66.5 -> 66.1 | Hurt |
| add multi-stage heatmap encoder without second-stage refinement (ablation) | center-based detector · nuScenes val | mAP: 63.1 -> 63.3 | No clear effect |
| add box-pooling RoI refinement on top of multi-stage heatmaps (ablation) | FocalFormer3D · nuScenes val | mAP: 63.3 -> 65.1 | Improved |
| use full box-level deformable decoder (box pooling plus cross attention) (ablation) | FocalFormer3D · nuScenes val | mAP: 65.1 -> 66.5 | Improved |
| add rescoring only without box regression (ablation) | FocalFormer3D · nuScenes val | mAP: 63.1 -> 66.1 | Improved |
| use box-level query embedding instead of point-level (ablation) | FocalFormer3D decoder · nuScenes val | mAP: 65.9 -> 66.5 | Improved |
| use deformable attention instead of full cross attention (ablation) | FocalFormer3D decoder · nuScenes val, 600 queries | latency: 26.2ms (full, 2 layers) vs 17.0ms (deformable, 6 layers) | Improved |
Later papers that compared with this method
4 later accepted papers report FocalFormer3D as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- BEVDilation: LiDAR-Centric Multi-Modal Fusion for 3D Object Detection (AAAI 2026 · better on 2 of 2 shared results)
- GAFusion: Adaptive Fusing LiDAR and Camera with Multiple Guidance for 3D Object Detection (CVPR 2024 · better on 1 of 1 shared result)
- UniM2AE: Multi-modal Masked Autoencoders with Unified 3D Representation for 3D Perception in Autonomous Driving (ECCV 2024 · better on 1 of 1 shared result)
- Diffusion Model for Robust Multi-Sensor Fusion in 3D Object Detection and BEV Segmentation (ECCV 2024 · better on 0 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- LiDARMultiNet — LidarMultiNet: Towards a Unified Multi-Task Network for LiDAR Perception (AAAI 2023) · better on 1 of 1 shared result
- TransFusion — TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection With Transformers (CVPR 2022) · better on 1 of 1 shared result
- DeepInteraction — DeepInteraction: 3D Object Detection via Modality Interaction (NeurIPS 2022) · better on 1 of 1 shared result
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
- Fully Sparse 3D Object Detection (NeurIPS 2022 · similarity 0.89)
- Super Sparse 3D Object Detection (TPAMI 2023 · similarity 0.89)
- DOPS: Learning to Detect 3D Objects and Predict Their 3D Shapes (CVPR 2020 · similarity 0.88)
- Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving (ICLR 2020 · similarity 0.88)
- NBA3D: Neighbor-Based Confidence Adjustment for 3D Rare Object Detection Using LiDAR (AAAI 2025 · similarity 0.88)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.87)
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.
