updates

Author: BinaryOracle

src/3DVL/LASO.md

@@ -281,28 +281,32 @@ class PointRefer(nn.Module):
         '''
 
         B, C, N = xyz.size()
-        # 使用RoBert编码文本 ，使用PointNet++编码点云
+        # 1. Encoding 过程
+        # 1.1 Language Encoding 使用RoBert编码文本
         t_feat, t_mask = self.forward_text(list(text), xyz.device)  # [batch, q_len, d_model]
+        # 1.2 BackBone Encoding 使用PointNet++编码点云
         F_p_wise = self.point_encoder(xyz)     
 
         """ 
         Decoding
         """
-        p_0, p_1, p_2, p_3 = F_p_wise  # 每个局部区域点坐标点，每个局部区域特征
-        p_3[1] = self.gpb(t_feat, p_3[1].transpose(-2, -1)).transpose(-2, -1) # 每个区域特征充分和文本信息进行融合
-        up_sample = self.fp3(p_2[0], p_3[0], p_2[1], p_3[1])   #[B, emb_dim, npoint_sa2] 特征传播
+        # 1.3 PointNet++ 逐级做点集抽象得到的每层的点集坐标和点集特征集合
+        p_0, p_1, p_2, p_3 = F_p_wise
+
+        # 2.Backbone Decoding过程
+        # 2.1 点集集合中每个点的特征和文本特征信息进行融合
+        p_3[1] = self.gpb(t_feat, p_3[1].transpose(-2, -1)).transpose(-2, -1)
+        # 2.2 PointNet++ 特征传播阶段: 上采样过程中，上一层点集中的点特征重建过程中，充分吸收了高级区域抽象特征和文本特征
+        up_sample = self.fp3(p_2[0], p_3[0], p_2[1], p_3[1])   #[B, emb_dim, npoint_sa2]
 
         up_sample = self.gpb(t_feat, up_sample.transpose(-2, -1)).transpose(-2, -1)
         up_sample = self.fp2(p_1[0], p_2[0], p_1[1], up_sample)    #[B, emb_dim, npoint_sa1]   
 
-        up_sample = self.gpb(t_feat, up_sample.transpose(-2, -1)).transpose(-2, -1)         
+        up_sample = self.gpb(t_feat, up_sample.transpose(-2, -1)).transpose(-2, -1) 
+        # 2.3 特征传播阶段结束: 一步步重建回原始点数量 128->256->512->1024->2048        
         up_sample = self.fp1(p_0[0], p_1[0], torch.cat([p_0[0], p_0[1]],1), up_sample)  #[B, emb_dim, N]
-
-        # t_feat = t_feat.sum(1)/(t_mask.float().sum(1).unsqueeze(-1))
-        # t_feat = t_feat.unsqueeze(1).repeat(1, self.n_groups,1)
-        # t_feat += self.pos1d
-
-        # print(t_feat.shape, up_sample.shape)
+        
+        #  3. Referred Point Decoding过程
         t_feat = self.decoder(t_feat, up_sample.transpose(-2, -1), tgt_key_padding_mask=t_mask, query_pos=self.pos1d) # b,l,c
         t_feat *= t_mask.unsqueeze(-1).float()
         _3daffordance = torch.einsum('blc,bcn->bln', t_feat, up_sample)