1. BEV感知

BEV Camera

view transformation

• 2d -> 3d via depth estimation
• 3d -> 2d (originates in 3d space)
• pure network based (implicitly)

BEV Lidar

voxelization and 3d convs

1. VoxelNet 2018(voxelization -> 3d convs -> flatten height dim -> RPN)

2. SECOND 2018(Sparse 3d convs)

No 3d convs (faster)

PointNet 2016(no voxels, using mlp encode pts features)

1. PointPillars 2019(Voxelization with pillars as BEV)

BEV Fusion

1. BEVFusion MIT(Efficient pv -> bev transformation)

temporal fusion

1. BEVDet4D 2022(Spatial alignment; concatenation of multiple feature map)

2. BEVFormer 2022(adopt a soft way to fusion temporal information)\(\star\)

1.1 dense bev feature

2. LSS 2020(First Depth Distribution)

3. BEVDet 2022(BEV space data augmentation)

4. BEVDepth 2022(Depth Correction)

5. BEVFusion 2022(Fusion on BEV from Camera and LiDAR)

Cam2BEV 2020(Homo-graphic Projection to BEV)

1.2 Attention Mechanism

1. DETR

2. DAB-DETR (收斂慢因為： 沒有提供位置先驗的 learnable queries)

3. DN-DETR (收斂慢因為：匈牙利匹配的離散性和模型訓練的隨機性，導致了 query 對 gt 的匹配變成了一個動態的、不穩定的過程)

4. Deformable DETR

5. DETR3D

6. PETR 2022(Implicit BEV Pos Embed)

7. BEVFormer 2022(Transformer for BEV feature)

Sparse3D

Recipe in practice

數據增強