search

hirotomusiker / CLRerNet

The official implementation of "CLRerNet: Improving Confidence of Lane Detection with LaneIoU"
Apache License 2.0
167 stars 18 forks source link

Custom dataset #41

Open okideal opened 3 months ago

okideal commented 3 months ago

Hey. Has anyone tried training with custom dataset? I manually labeled about 1000 images with lane, disabled crop-preprocess & segmentation loss, modified the max_lanes to 8(num_topk in CollectCLRNet & nms) The training loss can't converge in 50 epochs, and the image_demo outputs only the 4 weird lanes though my labels have 8.

okideal commented 3 months ago

specificly the loss_reg_xytl one, always larger than 50.0