Inference performance in KITTI dataset about second.pytorch

[traveller59]

I am not familiar with ROS, so I only can provide some advice.

1. There are some wrong boxes which are from previous frames, please check your code.
2. The pretrained only detect cars, if you want to detect vans you need to train a new model.
3. The pretrained is trained with the cars which is very close to self and is cropped during training. train with non-crop point cloud or crop the point cloud.
4. The second detector isn't so good in BEV benchmark, I think the major problem is noise per ground-truth. may be you need to fine-tune the network for several epoch without noise.
5. default config use low score threshold to get better kitti performance, consider increase it.

[cedricxie]

Thanks for the reply @traveller59 . I will follow your advice and try to improve the performance.

Thanks again for sharing your work!

Yuesong

[xieqi1996]

Hello,

Thanks @traveller59 for sharing the code! I tried to implement it as a ROS node repository link and test the performance with KITTI raw dataset 2011_09_26_drive_0005. You can find a video at the youtube link.

I suspect that I might have done something wrong and the performance could be improved. If any of you can check out my code and let me know if you have any suggestions / comments, feel free to do so. Thank you.

Best Regards,

Yuesong

Could you please tell me how to test the performance with KITTI raw dataset 2011_09_26_drive_0005? I don't know where to put the raw data.