RailEval

Deep Learning-based Rail Surface Condition Evaluation

Shilin Hu, Ke Ma, Sagnik Das, Dichang Zhang, Dimitris Samaras. ICCVW 2025. [link]

Accepted by the 3rd workshop on Vision-based InduStrial InspectiON(VISION).

Benchmark Dataset: TTC / Anomaly subset

Public access is pending FRA approval. The download link will be posted once approved.

For questions or early-access inquiries, contact shilhu@cs.stonybrook.edu.

Getting Started

conda env create -f environment.yml

The code is tested with python==3.7, torch==1.12.1, and CUDA ≥ 11.3.

Train

To train the model, run

python train_cls.py --dataroot ./cvdata_ttc/train_X.pkl --name TTC_X --checkpoints_dir $ckpt

Note: We use 4-fold cross-validation; the splits are provided in ./cvdata_ttc. Run the command for each of the four splits to reproduce the full results. The segmentation and alignment modules are pre-trained; update the paths in ./models/rail_newdata.py before training.

Evaluate

Our checkpoints are available at GoogleDrive

To test the model, run

python test_cls.py --dataroot ./cvdata_ttc/test_X.pkl --name TTC_X --checkpoints_dir $ckpt --results_dir $res_dir

Acknowledgements

This work was sponsored by the Federal Railroad Administration and a gift from NVIDIA Corp.
We also thank KLD Labs for their support and collaboration.

Miscellanous

Some parts of the code are based on @junyanz and @PRBonn.