This is a re-implementation of the keypoint network proposed in "Discovery of Latent 3D Keypoints via End-to-end Geometric Reasoning [pdf]". The keypointnet predicts a consistent set of keypoint given a single image. The predicted keypoint can then be used for various downstream tasks such as detection and pose estimation.
As seen in the images, the network is able to consistently detect the keypoint even with out of plane rotations.
Deepan Chakravarthi Padmanabhan
@inproceedings{suwajanakorn2018discovery,
title={Discovery of latent 3d keypoints via end-to-end geometric reasoning},
author={Suwajanakorn, Supasorn and Snavely, Noah and Tompson, Jonathan J and Norouzi, Mohammad},
booktitle={Advances in Neural Information Processing Systems},
pages={2059--2070},
year={2018}
}
The functions defined in this repository have been either adapted from or directly taken (Transformer class, blender render script and few of the loss functions) from https://github.com/tensorflow/models/tree/master/research/keypointnet following the license under the original repository.