Not the expected outcome when using a pre-trained model in ERFnet-CULane-pytorch

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

Not the expected outcome when using a pre-trained model in ERFnet-CULane-pytorch #143

Closed zhoufeng2 closed 5 years ago

zhoufeng2 commented 5 years ago

Thank you for you lane-detection code. When i test your pre-trained model in ERFnet-CULane-pytorch , i found the precision is not the expected result, the precision in the normal sence is just 35%, is there some parameters i should be adjusted?

cardwing commented 5 years ago

Do you use the ERFNet_trained.tar? It runs well in my local server. Edit: I have checked the trained model. The performance is exactly the same as the reported one. Using the testing scripts should work fine.

zhoufeng2 commented 5 years ago

yes, it is ERFNet_trained.tar. But the result is not the same as the reported one. Could it be because it's a single gpu in my local?

cardwing commented 5 years ago

I don't think so. The performance should be the same even if you use just one GPU. The following is the snapshot of the final performance I have just obtained.

73_1

zhoufeng2 commented 5 years ago

I am so confused that the results differ considerably. The following is the final preformance in my local:

anno_dir: ../../dataset/CULane/ detect_dir: ../../tools/prob2lines/output/vgg_SCNN_DULR_w9/ im_dir: ../../dataset/CULane/ list_im_file: ../../list/test.txt width_lane: 30 iou_threshold: 0.5 im_width: 1640 im_height: 590

Evaluating the results... tp: 45191 fp: 59005 fn: 17319 finished process file precision: 0.433711 recall: 0.72294 Fmeasure: 0.542164

cardwing commented 5 years ago

That's weird. I have not met this problem before. Please find out the cause by yourself.