Using segmentation masks as additional channels for classification
All implemented methods realised in separated branches of this repository to distinguish them and to ease code debugging.
- Base neural network based methods (NNmethods branch).
- Direct using masks methods:
- Zeroing masks (Zeroing_masks branch).
- Std loss (Aux_loss branch).
- Pair MSE loss (Pair_loss branch).
- Multi-task learning with Y-Net like architecture (Y_net branch).
- Future work on VOC dataset (VOC_ynet).
Each branch has description in its own Readme file.
All methods (except future work on VOC dataset) use the ISIC dataset. This dataset contains 2594 images of 5 melanocytic attributes: globules, milia-like cysts, negative network, pigment network, streaks&dots. Examples with topology of attributes you can see at the picture (from https://challenge2018.isic-archive.com/) below: