Neural network for detecting rails. NNs shows good result in semantic segmantation tasks.
- labeling dataset (89 images total) for NN to learn
- GPU was set up for speed up
- data agmentation for generating more data and improve model score
- used pretrained U-NET model from segmentation-models. It will help to converge faster.
Appropriate metric for this type of problem is IoU (intersection over Union),
since I care about matching pixel with rails.
Loss function is Jaccard index (it is the same as IoU, perfect function for this problem).
IoU score is about 0.80. Without any augmentation (used augmentation for more lighter model, but score was lower), and openCV tricks( more discussed about them in summary in notebook).
more information inside notebook