2016

[arxiv 1605.05180] Structured Prediction of 3D Human Pose with Deep Neural Networks [PDF] [notes]

Bugra Tekin, Isinsu Katircioglu, Mathieu Salzmann, Vincent Lepetit, Pascal Fua

Synthesis

Train an overcomplete encoder-decoder network and then use other CNN to predict the encoded vector, then decode.

The prediction + decoding structure is finetuned for pose estimation.

Autoencoder

Trained on a denoising task

Results

State of the art on Human3.6m