How do you visualize the BEV feature as in figure 1 of your paper?

[Jensen-Su]

I am working based on your method, but currently it did not go well as expected. To do debugging, I would like to check the step of converting img_feat to bev_feat via visualization. So how do you visualize the bev_feat as in the following figure of your paper?

[xuyanging]

Rewrite Conv2DCollapse and add following code: ' def forward(self, batch_dict): """ Collapses voxel features to BEV via concatenation and channel reduction Args: batch_dict: voxel_features [torch.Tensor(B, C, Z, Y, X)]: Voxel feature representation Returns: batch_dict: spatial_features [torch.Tensor(B, C, Y, X)]: BEV feature representation """ voxel_features = batch_dict["voxel_features"] # [bs, 64, 25, 376, 280] bev_features = voxel_features.flatten(start_dim=1, end_dim=2) # (B, C, Z, Y, X) -> (B, CZ, Y, X) bev_features = self.block(bev_features) # (B, CZ, Y, X) -> (B, C, Y, X) import cv2 bev_1 = bev_features[0][0].cpu().detach().numpy() for i in range(bev_features[0].shape[0]): bev = bevfeatures[0][i].cpu().detach().numpy() cv2.imwrite('test'+str(i)+'.jpg', bev*255/bev.max()) batch_dict["spatial_features"] = bev_features return batch_dict '

[Jensen-Su]

Rewrite Conv2DCollapse and add following code: ' def forward(self, batch_dict): """ Collapses voxel features to BEV via concatenation and channel reduction Args: batch_dict: voxel_features [torch.Tensor(B, C, Z, Y, X)]: Voxel feature representation Returns: batch_dict: spatial_features [torch.Tensor(B, C, Y, X)]: BEV feature representation """ voxel_features = batch_dict["voxel_features"] # [bs, 64, 25, 376, 280] bev_features = voxel_features.flatten(start_dim=1, end_dim=2) # (B, C, Z, Y, X) -> (B, C_Z, Y, X) bev_features = self.block(bev_features) # (B, C_Z, Y, X) -> (B, C, Y, X) import cv2 bev_1 = bev_features[0][0].cpu().detach().numpy() for i in range(bev_features[0].shape[0]): bev = bevfeatures[0][i].cpu().detach().numpy() cv2.imwrite('test'+str(i)+'.jpg', bev*255/bev.max()) batch_dict["spatial_features"] = bev_features return batch_dict '

Your meaning, just to visualize each channel of the collapsed feature map (B, C, Y, X)？

[xuyanging]

Rewrite Conv2DCollapse and add following code: ' def forward(self, batch_dict): """ Collapses voxel features to BEV via concatenation and channel reduction Args: batch_dict: voxel_features [torch.Tensor(B, C, Z, Y, X)]: Voxel feature representation Returns: batch_dict: spatial_features [torch.Tensor(B, C, Y, X)]: BEV feature representation """ voxel_features = batch_dict["voxel_features"] # [bs, 64, 25, 376, 280] bev_features = voxel_features.flatten(start_dim=1, end_dim=2) # (B, C, Z, Y, X) -> (B, C_Z, Y, X) bev_features = self.block(bev_features) # (B, C_Z, Y, X) -> (B, C, Y, X) import cv2 bev_1 = bev_features[0][0].cpu().detach().numpy() for i in range(bev_features[0].shape[0]): bev = bevfeatures[0][i].cpu().detach().numpy() cv2.imwrite('test'+str(i)+'.jpg', bev*255/bev.max()) batch_dict["spatial_features"] = bev_features return batch_dict '

Your meaning, just to visualize each channel of the collapsed feature map (B, C, Y, X)？

That is the BEV feature you refered to in your question.