BEVFormer Code Analysis

(mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py)

class BEVFormerHead()

 

(mmdet3d_plugin/modules/bevformer/transformer.py)

class PerceptionTransformer()

 

(mmdet3d_plugin/modules/bevformer/encoder.py)

class BEVFormerEncoder()

class BEVFormerLayer()

 

(mmdet3d_plugin/modules/bevformer/temporal_self_attention.py)

class TemporalSelfAttention()

 

(mmdet3d_plugin/modules/bevformer/spatial_cross_attention.py)

class SpatialCrossAttention()

class DeformableAttention3D()

 

TODO List (Reproducing purpose)

1) Use key [Q, Q] instead of [Q] in 'SpatialCrossAttention'

2) Integrate 'z anchors' into deformable attention (DeformableAttention3D())

3) Reproduce attention decoder

4) Train model with proposed hyper-parameters

5) Train model with proposed loss functions (Impossible)