search

pangsu0613 / CLOCs

CLOCs: Camera-LiDAR Object Candidates Fusion for 3D Object Detection
MIT License
351 stars 68 forks source link

Is the SECOND code in your project the same as that of the original author? #65

Closed xuV2 closed 2 years ago

xuV2 commented 2 years ago

If so, can I directly look at the code you wrote without knowing the SECOND code?

xuV2 commented 2 years ago

@pangsu0613

pangsu0613 commented 2 years ago

Hello @xuV2, thank you for your interests in CLOCs. It is a little bit hard to say, I did make modifications in the original SECOND code, but not the core model part. But if you have any questions regarding the code, please feel free to ask.

xuV2 commented 2 years ago

Hello @xuV2, thank you for your interests in CLOCs. It is a little bit hard to say, I did make modifications in the original SECOND code, but not the core model part. But if you have any questions regarding the code, please feel free to ask.

Thank you for your reply！

xuV2 commented 2 years ago

Hi @pangsu0613 I want to ask you two questions, 1.What is self.corner_points_feature in fusion.py ？ 2.Is the tensor_index with flag=0 in fusion.py finally discarded ？Thank you.

pangsu0613 commented 2 years ago
  1. the corner_points_feature is not used. I should have deleted it, sorry for the confusion. I used it to test with fusing with boounding boxes corner points coordinates, but the results turned to be not good, so I ended up with not using it.
  2. tensor_index with flag = 0 means there are no overlaps between the 3D detections and 2D detections in the current frame, so we cannot learn anything, so they are not contribute in the training, or say discarded.
xuV2 commented 2 years ago

Excuse me again @pangsu0613 ，I would appreciate it if you could explain the code all_3d_output_camera_dict, all_3d_output, top_predictions, fusion_input,tensor_index = net(example_torch,detection_2d_path) in train.py.