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:
- we have labelled data for domain A(from such as public dataset, or synthetic data)
- we have unlabelled data for domain B (the domain is where you want to deploy your algorithm)
- there are some shift or difference between domain A and domain B ( weather condition, scenario difference)
- the camare intrinsic parameter for both domain is available
- Python
- Pytorch numpy scipy pandas scikit-image
- GPU
- VO Feature Extraction
- VO regression
- Discriminator
- 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)