search

hirotomusiker / CLRerNet

The official implementation of "CLRerNet: Improving Confidence of Lane Detection with LaneIoU"
Apache License 2.0
181 stars 19 forks source link

How to integrate the LaneIoU loss function from CLRerNet into the LaneATT architecture? #48

Closed XiaoCangbit closed 1 month ago

XiaoCangbit commented 1 month ago

Hello! Could you kindly assist me in integrating the LaneIoU loss function from CLRerNet into the LaneATT architecture? Additionally, could you guide me on where I might find the relevant code? Thank you in advance for your help.

hirotomusiker commented 1 month ago

Thank you.

We described an effective LaneATT + LaneIoU setting in the supplementary material.

Positive and negative anchors are calculated separately for classification loss and regression loss, and LaneIoU is employed only for the former. The anchors whose max IoU with GTs is over 0.7 (under 0.5) are regarded as positive (negative) and the classification target is set to 1 (0). (ref: LaneATT matching function)

We don't see significant improvement when LaneIoU is applied for regression loss.

hirotomusiker commented 1 month ago

We will publish LaneIoU-aware match_proposals_with_targets function later for reference.