search

HorizonRobotics / Sparse4D

MIT License
326 stars 31 forks source link

about distant target #77

Open AWDawson opened 3 months ago

AWDawson commented 3 months ago

If there are distant targets in the dataset that are visible in the images but are not included in the ground truth, this situation could harm model training. How can this issue be addressed?

linxuewu commented 3 months ago

Limit the perception range of the model to be consistent with the ground truth (GT) range to minimize the impact of GT missed detections. If the GT missed detections are severe, there's not much that can be done. You can try to reduce the classification loss weight for distant negative samples, although I haven't tried that myself.