search

cardwing / Codes-for-Lane-Detection

Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)
MIT License
1.04k stars 333 forks source link

Some questiones on tensorflow version on CULane #312

Closed hustcc19860606 closed 2 years ago

hustcc19860606 commented 3 years ago

Following the README， i can get:

anno_dir: /home/cc/SCNN-Tensorflow/data/CULane/ detect_dir: /home/cc/SCNN-Tensorflow/data/CULane/result-matlab/scnncastedBGR/vgg_SCNN_DULR_w9/data/CULane/ im_dir: /home/cc/SCNN-Tensorflow/data/CULane/ list_im_file: /home/cc/SCNN-Tensorflow/data/CULane/test.txt width_lane: 30 iou_threshold: 0.5 im_width: 1640 im_height: 590 Evaluating the results... tp: 76970 fp: 34736 fn: 27916 finished process file precision: 0.689041 recall: 0.733844 Fmeasure: 0.710737

My questiones have two:

1.The gap between 0.710737 and 71.3 i s caused by tensorflow? 2.If i want to reproduce the result by a new training, do i need to modify the VGG_MEAN = [123.68, 116.779, 103.939] into VGG_MEAN = [103.939, 116.779, 123.68] in lanenet_data_processor.py?

@cardwing

cardwing commented 3 years ago
  1. Maybe yes. You can refer to previous issues for more details.
  2. You can test both variants.
hustcc19860606 commented 3 years ago

From question 81#，i find“My original model can achieve around 69.5 when training from scratch.” from your answer. However, i only get 68.164 on F1-measure. And the training takes 90000 epoches as the your setting. Do you take some tricks on the original model? @cardwing