Anyone reproduce the results in the paper?

[RaymondByc]

Thanks for the great work!

I am reproducing the whole project, but I found the reproduced results are far from the expected, and I have some questions.

1. How to train the detector? Could you provide a detailed script as the paper described?
2. The detector is coupled with the tracking parameters? I converted my detection results with the required format, but the tracking results lag behind the expected results.
3. Why the GRM module does not use the sigmoid to force the final result larger than zero? I found this problem because my GRM results have some refined size smaller than zero.

[RaymondByc]

I have another questions about CRM part.

The input of CRM is the point cloud cropped by the unrefined boxes, but the expected result iou is computed by the refined boxes. The gap between the unrefined boxes and the refined boxes will confuse the CRM module.

Why not reproduced the point cloud by the new boxes?