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
785 stars 100 forks source link

How to improve the robustness of fsdv2 algo？ #153

Closed eagle-chase closed 1 year ago

eagle-chase commented 1 year ago

In the FSDv2 algorithm, distant targets may have very few points in the point cloud. If these points are classified as background, it can result in missed detections. How can we improve the detection accuracy of distant targets? For close-range targets, false positives are often encountered, such as mistaking objects like guardrails、trees for vehicles. Is there a way to reduce false positives and missed detections in algorithm design? Does the occurrence of these false positives and missed detections primarily depend on the accuracy of the segmentation network?

eagle-chase commented 1 year ago

I also want to know if it is a feasible method to filter the ground truth boxes based on the number of point clouds in the ground truth boxes, such as filtering out boxes with less than 5 point clouds.

eagle-chase commented 1 year ago

If I want to increase the depth of the model, which part of the model do you think would be best to modify?

Abyssaledge commented 1 year ago
  1. Segmentation theoretically has an influence on detection accuracy. However, it seems fine in FSDv2. If a distance object is completely classified as background, then I guess it is also difficult for other detectors to detect the object.
  2. Of course you can filter these boxes. Just calculate the number of contained points. It seems straightforward.
  3. It is tricky. I think the current backbone is large enough.