this code is based on https://github.com/corenel/pytorch-adda trying to investigate the domain adaption for learning based visual odometry problem.

The assumption of this world is that:

1. we have labelled data for domain A(from such as public dataset, or synthetic data)
2. we have unlabelled data for domain B (the domain is where you want to deploy your algorithm)
3. there are some shift or difference between domain A and domain B ( weather condition, scenario difference)
4. the camare intrinsic parameter for both domain is available

• Python
  • Pytorch numpy scipy pandas scikit-image
• GPU

• VO Feature Extraction
• VO regression
• Discriminator

1. intrinsic remapping

We proposed to train a deep visual odometry model with a pure random dataset created by following procedure:

image_1 = np.random.random((1,self.camera_parameter[1],self.camera_parameter[0]))

depth   = np.random.random((1,self.camera_parameter[1],self.camera_parameter[0]))

motion  = np.random.random((6))   
image_2 = remap(image_1,depth,motion)