alldbi / Factorized-Spatial-Embeddings

Tensorflow implementation of Unsupervised learning of object landmarks by factorized spatial embeddings

Factorized-Spatial-Embeddings

Tensorflow implementation of Unsupervised learning of object landmarks by factorized spatial embeddings by Thewlis el al. for unsupervised landmark detection.

Sample results

Test results on LFW with 8 landmarks (K=8, M=4), trained on CelebA dataset for 2 epochs. Test results on LFW with 16 landmarks (K=16, M=4), trained on CelebA dataset for 2 epochs.

Setup

Prerequisites

• Tensorflow 1.4

Getting Started

First download the CelebA dataset or the UT Zappos50k shoes dataset, extract images and use them to train the model.

# clone this repo
https://github.com/alldbi/Factorized-Spatial-Embeddings.git
cd Factorized-Spatial-Embeddings
# train the model 
python main.py \
  --mode train \
  --input_dir (directory containing CelebA dataset) \ 
  --K 8  \ #number of landmarks to be learned

# test the model
python main.py \
  --mode test \
  --input_dir (directory containing testing images)
  --checkpoint (address of the trained model, which is /OUTPUT as default)
  --K 8