Feature Dynamic Aggregation Module

[hbsycym]

hello，I have a question that where the code for Feature Dynamic Aggregation Module is ? I have seen the code in LoGoHead_kitti.py

pooled_features = pooled_features + localgrid_densityfeat_fuse.permute(0, 2, 1) and attention_output = pooled_features + attention_output

if this code means the three feature F p B + F l B + F g B, then I didn't see the self attetion moudle before

shared_features = self.shared_fc_layer(pooled_features.view(batch_size_rcnn, -1, 1))

[sankin97]

Please refer to line557 in LoGoHead_kitti.py and line134 in attention_utils.py.

[hbsycym]

Please refer to line557 in LoGoHead_kitti.py and line134 in attention_utils.py.

thanks ！ and one more question， the input_sp_tensor of spconv.SparseConvTensor in backbone3d, its spatial shape is [41, 1600, 1408] , is it mean how many voxels on Z, Y, X axis respectively ？

[1806610292]

你好，我有个问题，特征动态聚合模块的代码在哪里？ 我在 LoGoHead_kitti.py 中看到了代码

pooled_features = pooled_features + localgrid_densityfeat_fuse.permute(0, 2, 1) 和tention_output = pooled_features +tention_output

如果此代码表示三个特征 F p B + F l B + F g B，那么我之前没有看到自我注意模块

shared_features = self.shared_fc_layer（pooled_features.view（batch_size_rcnn， -1， 1））

你好，我没看懂作者的意思，现在我在代码里没有对应找到LoF/GoF/FDA三个模块的代码，attention_output = self.attention_head(pooled_features, positional_input, src_key_padding_mask) 这句代码的描述的是GoF还是FDA，期待你的回答，谢谢！