GANVO

Generative Adversarial Network for Visual Odometry Pytorch Implementation of the paper -- GANVO

Disclaimer: This is not an official release. This implementation is based on the ICRA 2019 paper of the same title by Yasin Almalioglu1 ,Muhamad Risqi U. Saputra, Pedro P. B. de Gusmo, Andrew Markham, and Niki Trigoni. I am trying to reproduce the results, while incorporating my own interpretations of the approach, wherever needed. Please check back here for the detailed results!

The work uses GANs for unsupervised visual odometry which is later used for depth estimation.

Training

• Prepare the config file -- this specifies the various hyperparameters, weights directory, summary, the root directory of the dataset, etc.
• Install all the dependencies with pip install -r requirements.txt
• Run python train.py --config ./config.yaml
• For pretrained weights visit this link

Testing

• Testing Pose -- run python eval_pose.py --config ./eval_pose.py for quantitative pose evaluation.

Inference

• For depth inference on a single image, run python inference.py --img_path ./test_image.png --generator_weights ./weights/generator.pth --h 128 --w 416

Qualitative Results

update very soon

Quantitative Results

update very soon