search

zhouhuan-hust / LFD_RoadSeg

[IEEE TITS] Exploiting Low-level Representations for Ultra-Fast Road Segmentation
MIT License
18 stars 1 forks source link

Questions Regarding Validation (val) and Five-Fold Cross-Validation #2

Closed qizaiCSDN closed 7 months ago

qizaiCSDN commented 7 months ago

Hello author, first of all, thank you for sharing your excellent work on GitHub! When I was reproducing your code, I encountered a problem in the validation (val) part. After running the code, an error occurred, indicating that there is no val_loader parameter in line 84 of _LFDRoadSeg.py. After checking, I found that in the ValDataset class in kitti_dataloader.py, the data obtained is from the kitti/testing directory, and then the labels are retrieved from the gt_image_2 folder under testing. I believe this is what caused the error, because in the downloaded KITTI dataset and the dataset structure you provided in the README, there is no gt_image_2 under testing. I tried to change the dataset root directory to test under exp, but I still encountered other issues. How should I modify it?

Another question is about five-fold cross-validation. When I was building KITTIVAL, I used a Python script to randomly split the training data into five parts, with one part placed in test under exp and the other four parts placed in train. I repeated this process five times to create exp1 to exp5. I'm not sure if my approach is consistent with your idea of building KITTIVAL. Could you please tell me which Python file contains the specific code implementation for your five-fold cross-validation? I would like to check my approach based on your code.

I apologize for the many questions, but I would greatly appreciate your answers! Thank you very much!

zhouhuan-hust commented 7 months ago

Hello, thank you for your interest in my work on GitHub. Regarding the validation part in the code, it is used to evaluate the performance metrics on the test data for each fold during five-fold cross-validation. There are two steps involved: constructing the kittival dataset and modifying the parameter configurations in the configs file.

For constructing the kittival dataset, I divided the images based on their sequential numbering. Specifically, the test set for exp1 includes images with sequential numbers 000000, 000005, 000010, etc., while the other four sets are used for training. Similarly, for exp2, the test set includes images with numbers 000001, 000006, 000011, etc., and so on for exp3, exp4, and exp5.

Regarding the modification of parameter configurations in the configs file, you need to set the dataset_name to "kittival" and the data_path to /xxx/data/kittival/exp1. You can also adjust the outputs path as needed.

I hope my explanation helps you. If you have any further questions, feel free to ask!

qizaiCSDN commented 7 months ago

Thank you for your patient answer! I have solved the problem. Thank you very much!