Sparse Fusion Detectors

fully sparse
1. SparseLIF, 24
  1. (mAP and NDS on nuScenes: 74, 77)
  2. (mAP on AV2 val set: 40)
  3. 它的实验中，比**EA-LSS 23好。**
2. SparseFusion, 24
  1. (mAP and NDS on nuScenes: 70.1, 72.7)
  2. (mAP on AV2 val set: 39.8)
3. SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection, iccv23
4. Futr3d: A unified sensor fusion framework for 3d detection, cvpr23
5. Deepinteraction: 3d object detection via modality interaction, nips22
sparse + global attention
1. Cross Modal Transformer, iccv23 (mAP and NDS on nuScenes, test: 72, 74)
2. UniTR: A unified and efficient multi-modal transformer for bird’s-eye-view representation, iccv23
3. exhaustive global attention buries the advantages of the sparse paradigm and makes it difficult to benefit from long-term temporal information
instance based
1. Fully Sparse Fusion for 3D Object Detection, TPAMI24
  1. (mAP and NDS on nuScenes: 70, 74)
  2. (mAP on AV2 val set: 33)
2. ObjectFusion, iccv23 (mAP and NDS on nuScenes: 71, 74)
3. PoIFusion, 24 ((mAP and NDS on nuScenes: 74, 75)
Other
1. **EA-LSS 23:** (mAP and NDS on nuScenes: 77, 78)
  1. EA-LSS [20] enhances depth estimation at the edge of objects