- visual odoemtry calculation does not rely on the whole image. A featureful sub-region (patch) is enough for estimating an ego-motion.
- The images patches have more similarity over scene.
- Estimation from patches can reduce the estiamtion variance
- Estimation from patches can increase the training data amount implicitly
- Estimated a ego-motion from each patch together with a reliability
- Achieve the overall ego-motion and reliability
- Coming soon
- only if you can build this docker image had have at least 8G GPU
- install all the requirement as mentioned in the dockerfile.
cd dockerfile
docker build -t xxx/pytorch .
- run the docker container
- create the dataset folder for your dataset like kitti folder there;
- modify script train.sh about for some option, for the meaning for options refer to src/option.py
- run the code in the docker container
sh script/train.sh
