Open sunwonHong opened 3 years ago
LiDAR-based or RGB-D-based object detection is used in numerous applications, ranging from autonomous driving to robot vision. Voxel-based 3D convolutional networks have been used for some time to enhance the retention of information when processing point cloud LiDAR data. However, problems remain, including a slow inference speed and low orientation estimation performance. We therefore investigate an improved sparse convolution method for such networks, which significantly increases the speed of both training and inference. We also introduce a new form of angle loss regression to improve the orientation estimation performance and a new data augmentation approach that can enhance the convergence speed and performance. The proposed network produces state-of-the-art results on the KITTI 3D object detection benchmarks while maintaining a fast inference speed. View Full-Text
this is abstract of paper about second:sparsely convolution detection. I think second is not a name of network, second is a name of method for reducing the operating time when use voxelnet.
Thank you so much for your kind response. 감사합니다.
I understand it is like a method for Detection. but i want to know whether it is different or not with pointpillars Net code below. https://github.com/nutonomy/second.pytorch
Could you let me know the difference?
If what I know is right, pointpillars is not a network, is a method
if you use config file in second.pytorch/second/config/pointpillars/ you can also pointpillars method with second method.
before you train, you must install spconv, is right?
spconv`s full name is sparsely convolution spconv make you can use sparsely convolution(second method)
and when you use config file in folder of pointpillar, you can also use pointpillars method with second method
The coment is what I understood while studying, and there may be a misunderstanding.
and I want advice about evaluation, In previous issue I think I saw that you complete test with costom dataset at first glance.
I struggle to evaluate second.pytorch with custom dataset.
I want that second.pytorch evaluate with custom dataset, and I want to also visualize this evaluation. do you complete this process? can you give some advice for me
my e-mail address is sjkalatey@naver.com.
If you have advice, please give me mail about advice.
And I hope that ,if you can use korean, advice mail consist of korean
Thank you so much. So, in order to use pointpillars method, you said that we should use configs in pointpillars folder. And it means Car.fhd.config and All.fhd.config is not for Pointpillar method. isn't it?
Additionally for custom dataset, Let me e-mail to you :D
Thank you so much. So, in order to use pointpillars method, you said that we should use configs in pointpillars folder. And it means Car.fhd.config and All.fhd.config is not for Pointpillar method. isn't it?
Yes (if what I understand is right)
is this code's first setting Second Network or Point Pillar Network?
If it is not Point Pillar Network, how can i change for that? and what is the difference?