search

zillur-av / LVLane

This repository implements a lane detection and classification model.
Other
20 stars 1 forks source link

Training problem about my own dataset #7

Closed xiaoxiaoyi689 closed 10 months ago

xiaoxiaoyi689 commented 1 year ago

Hi author，thanks for your great work!

I have created my dataset which is same form as Tusimple dataset. My image size is 1080x1920，and I have changed the image size from the files "tools/generate_seg_tusimple.py" and "configs/ufld/resnet18_tusimple.py". But when training, the ''classification accuracy'' keeps increasing to about 94%, while "detection metric" grows slowly and stops growing at about 39%. The test results are as follows; 截图 2023-11-10 15-42-07

xiaoxiaoyi689 commented 1 year ago

I don't know if this problem has anything to do with the size of the my dataset, and my training set is only 500 images

zillur-av commented 1 year ago

It has nothing to do with dataset size but the resolution of images. The best way to handle this would be resize your images to 1280x720 and modify annotations proportionately as well. Another quick step you can take is to modify these lines https://github.com/zillur-av/LVLane/blob/f89d53d63b45069fdae6689157c7f33caa6c8652/configs/ufld/resnet18_tusimple.py#L56-L61

Change orig_img_h, orig_img_w, sample_y. For example, if your images are 1920x1080, you can change sample_y to range(1070, 450, -10) something like that. It depends on how many lane points you have in your training images. Like I had 56 points from 150 to 710 and I used 56 here https://github.com/zillur-av/LVLane/blob/f89d53d63b45069fdae6689157c7f33caa6c8652/configs/ufld/resnet18_tusimple.py#L20-L22 Probably you need to change that as well. Check your json labels to know from where you start annotating your lane points. Just give it a shot and let me know.

xiaoxiaoyi689 commented 1 year ago

Thank you for your reply. I will tell you the experimental results as soon as possible

xiaoxiaoyi689 commented 1 year ago

Hi, author, I have solved the above problem. I scaled the image to 1280x720 , and modify the annotations. Finally, both the training and test results looked normal. Thanks again for your help.

gabaebae123 commented 4 weeks ago

Hi, author, I have solved the above problem. I scaled the image to 1280x720 , and modify the annotations. Finally, both the training and test results looked normal. Thanks again for your help.

Could you tell me how you make your own data set?