Research map / Detection and segmentation / 3D detection and driving
Make Your ViT-based Multi-view 3D Detectors Faster via Token Compression
Dingyuan Zhang, Dingkang Liang, Zichang Tan, Xiaoqing Ye, Cheng Zhang, Jingdong Wang, 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 |
|---|---|---|---|
| replace random token selection with motion-query-guided selection (ablation) | ToC3D on StreamPETR · nuScenes val (ViT-L, 12 epochs) | NDS (%): 48.5 -> 60.3 | Improved |
| use motion query-guided selection instead of DynamicViT / SparseDETR (ablation) | ToC3D · nuScenes val (ViT-L, 12 epochs) | mAP (%): 49.3 -> 51.2 (vs DynamicViT) | Improved |
| compute importance score via attention between tokens and history queries (ablation) | ToC3D-Faster · nuScenes val (ViT-L) | mAP (%): 50.4 -> 51.2 | Improved |
| add motion vector encoding of history queries (ablation) | ToC3D-Faster · nuScenes val (ViT-L) | mAP (%): 50.1 -> 51.2 | Improved |
| sample top-confidence history queries instead of using all (ablation) | ToC3D-Faster · nuScenes val (ViT-L) | mAP (%): 50.4 -> 51.2 | Improved |
| add bridge token in dynamic router (ablation) | ToC3D-Faster · nuScenes val (ViT-L) | mAP (%): 50.5 -> 51.2 | Improved |
| use too few history queries (N_q too small) (ablation) | MQTS · nuScenes val | performance | Hurt |
| keeping ratio too small (0.3, 0.2, 0.1) (ablation) | ToC3D · nuScenes val | NDS/mAP | Hurt |
| update importance score at deeper layers (10, 16, 22 vs 6, 12, 18) (ablation) | ToC3D-Faster · nuScenes val (ViT-L) | NDS (%): 60.3 -> 61.0 NDS; 237.2 -> 264.0 ms | Mixed |
| apply ToC3D to Sparse4Dv2 as plug-and-play | Sparse4Dv2 · nuScenes val (ViT-L) | NDS/speed: NDS 58.8 -> 58.8; 322.0 -> 244.8 ms | Improved |
| apply ToC3D-Fast to StreamPETR ViT-B | StreamPETR ViT-B · nuScenes val | whole pipeline time (ms): 115.0 -> 94.0 with NDS 55.4 -> 55.2 | Improved |
Closest papers by meaning
- Revisiting Token Compression for Accelerating ViT-based Sparse Multi-View 3D Object Detectors (CVPR 2026 · similarity 0.92)
- LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging (CVPR 2026 · similarity 0.87)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.85)
- Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers (NeurIPS 2026 · similarity 0.85)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.85)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.85)
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.
