SharinGAN

Official repo for the work titled "SharinGAN: Combining Synthetic and Real Data for Unsupervised GeometryEstimation"

The official project website for this work can be found here.

Requirements

Python 2.7 Pytorch 0.4.1

The trained model files for both the tasks are made available here here. The pretrained_models folder contains the pretrained model for the generator and the primary task networks before end-to-end training the SharinGAN network as a whole. The final trained models are present in Face_Normal_Estimation/ and Monocular_Depth_Estimation/ directories of the google drive.

The environment.yml file is also provided for one to replicate the environment.

We added the training and validation codes for both the tasks of Monocular Depth Estimation and Face Normal Estimation. We hope to improve the repository with time. We appreciate your inputs and feedback

Monocular Depth Estimation

Place the saved model file (Depth_Estimator_WI_geom_bicubic_da-144999.pth.tar) inside a newly created folder Monocular_Depth_Estimation/saved_models/ of the current repo.

The dataset files required for the dataloaders Kitti_dataloader.py and VKitti_dataloader.py are made available at Monocular_Depth_Estimation/dataset_files/.

Place the Monocular_Depth_Estimation/dataset_files/Kitti/.txt files in the original downloaded kitti/ dataset folder. Similarly place the Monocular_Depth_Estimation/dataset_files/VKitti/.txt files in the original downloaded Virtual_Kitti/ dataset folder.

Make3D evaluation

cd Monocular_Depth_Estimation
python Make3D_validation.py --iter 144999