search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
765 stars 96 forks source link

Using the Ground truth boxes as a history offline predictions in FSD++ #178

Open rXYZkit opened 6 months ago

rXYZkit commented 6 months ago

Thank you very much for your work.

Have you tried using the Ground truth boxes as a history offline predictions versus using the predictions of a well-trained fsd model as a historical prediction even if you mention in your paper that the distribution gap between ground truth used in training and predicted boxes.

Abyssaledge commented 6 months ago

No, I have not tried it. I believe there will be a performance gap if using the GT, which may not be very significant due to the well-trained FSD having high-quality predictions. Moreover, you can narrow the gap by randomly jittering the GT.

rXYZkit commented 6 months ago

Thanks,

The predictions of a trained model will still contain FP and TN, which bboxes are not as accurate as gt produces for generating historical skeleton points when training. You mentioned using random jittering for gt, is that because less accurate skeleton points are more friendly to training?