TimingSpace / PADVO

Patch Agreement Deep Visual Odoemtry

PAD-VO: Patch Agreement Deep Visual Odometry

Motivation:

1. visual odoemtry calculation does not rely on the whole image. A featureful sub-region (patch) is enough for estimating an ego-motion.
2. The images patches have more similarity over scene.
3. Estimation from patches can reduce the estiamtion variance
4. Estimation from patches can increase the training data amount implicitly

Method

1. Estimated a ego-motion from each patch together with a reliability
2. Achieve the overall ego-motion and reliability

Paper

1. Coming soon

To Run This Code:

Requirement

1. only if you can build this docker image had have at least 8G GPU
2. install all the requirement as mentioned in the dockerfile.

cd dockerfile
docker build -t xxx/pytorch .

Run

1. run the docker container
2. create the dataset folder for your dataset like kitti folder there;
3. modify script train.sh about for some option, for the meaning for options refer to src/option.py
4. run the code in the docker container

sh script/train.sh

Interesting discovery