Unsupervised Domain Adaptation for Semantic Segmentation of NIR Images through Generative Latent Search

Tensrflow Implementation of our method to transfer the bias from the source to the target domain. The latent space of a generative model, learnt on the source domain, is employed to find the "clones" for the target samples. As the "clones" are sampled from the source distribution, an oracle segmentation model learnt only on source, will interpret these "clones" of target samples in a better way, thereby reducing the domain shift. The proposed method guarantees to reduce the shift.

Paper

'Unsupervised Domain Adaptation for Semantic Segmentation of NIR Images through Generative Latent Search' Accepted in ECCV2020 (Spotlight). Preprint: https://arxiv.org/abs/2006.08696

Example Results

Quantitative Results

Dataset Access

Kindly fill the form to download the dataset https://forms.gle/y5vPeyT6zi9gdqD57 On filling the form, the datasets namely SNV dataset and Hand Gesture dataset will be shared by email.

Requirements

tensorflow = 1.14.0
python = 3.6
keras = 2.2.5

Preprocessing

Image were reshaped to 128x128 dimension and normalized between -1 to 1.

Training

Train segmentation network on the source domain

python u_efficientnet.py

To train VAE and Perfom Latent search on Target domain.

python vae_u-efficientnet_latent_search.py

To Train Vae with Perceptual loss with Deeplabs

python vae_DeepLabs-frez--PL-LS.py

Cite

@article{pandey2020skin,
  title={Skin Segmentation from NIR Images using Unsupervised Domain Adaptation through Generative Latent Search},
  author={Pandey, Prashant and Tyagi, Aayush Kumar and Ambekar, Sameer and AP, Prathosh},
  journal={arXiv preprint arXiv:2006.08696},
  year={2020}
}

ambekarsameer96/GLSS