Learning transformations between vision datasets to overcome sensor hardware versioning
paper: IEEE Robotics and Automation Letters (RA-L) video: YouTube
The main idea of this project is to come up with a universal transformation between dataset distributions. The difference between these dataset distributions are a result of hardware versioning between sensors. This transformation should preserve several properties of the original dataset:
- The prediction of the base model.
- The structure and features present in the image.
Problem setup
Input and output of trained PreFixer
pip install -r requirements.txt
cd training
python train_lenscoder.py <path-to-dataset>