This is an unofficial PyTorch implementation of the paper:
GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose (CVPR 2018)
Zhichao Yin and Jianping Shi
arxiv preprint: (https://arxiv.org/abs/1803.02276)
The official tensorflow implementation: https://github.com/yzcjtr/GeoNet
Build on:
python 3.7
PyTorch 1.0 stable
CUDA 9.0
Ubuntu 16.04 / CentOS 7
This code follows the GeoNet authors stage-wise training as:
- "train_rigid" mode: Train DispNet and PoseNe with rigid warp loss, smooth loss
- "train_flow" mode: Fine tune/fix DispNet and PoseNet and train FlowNet with rigid warp loss, rigid smooth loss, fully warp loss, fully flow smooth loss and geometry consistency loss
This repository is still work on progress and here are the todos:
- validation functions with ground truth
- test functions for depth, pose and optical flow
- evaluation
- debugging and experiments
This repository heavily reused codes from SfmLearner-Pytorch by Clement Pinard. Many thanks to Clement Pinard's work!