Simlearner3D > Documentation
Simlearner3D is a deep learning library designed with the focus of large scale dense image matching by similarity learning from pairs of epipolar images.
The library implements the training of feature extractors (MSAFF,UNet32,UNet-Attention) with and without a MLP given pairs of images, corresponding disparity and occlusion maps. Qualification of models is conducted using joint probability maps estimation. This actually tells if learned similarity cues allow decent separation between matching and non matching pixels.
Simlearner3D is built upon PyTorch. It keeps the standard data format from Pytorch-Geometric. Its structure was bootstraped from this code template, The latter relies on Hydra and Pytorch-Lightning to enable flexible and rapid iterations of deep learning experiments.
Getting Started