Deep Unsupervised Similarity Learning Using Partially Ordered Sets (CVPR 2017)

Accepted at CVPR 2017
"Deep Unsupervised Similarity Learning Using Partially Ordered Sets" Miguel A. Bautista* , Artsiom Sanakoyeu* , Björn Ommer.

• Paper: https://arxiv.org/abs/1704.02268
• GT labels for Olympic Sports dataset: olympic_sports_retrieval/data
• Evaluation script for Olympic Sports dataset: calculate_roc_auc.py
• Baseline HOG-LDA similarity matrices for Olympic Sports: similarities_hog_lda.tar.zip (11.5 Gb)

Tensorflow models for Olympic Sports dataset trained with our approach

All models were trained from scratch without Imagenet pretraining and without any supervision.

1. The model trained on all frames from Olympic sports dataset: olympic_sports_all_cat_convnet_scratch_strip.ckpt
2. Using the same method we finetuned the previous model for each sport independently w/o any supervision (we again used only grouping and posets that we build without GT information).
   Single models for each sport: olympic_sports_models_from_scratch

Requirements

• Python 2.7
• Tensorflow r1.*

Example

Example how to load models: example_load_networks.ipynb.

If you find this code or data useful for your research, please cite

@inproceedings{UnsupSimPosets2017,
  title={Deep Unsupervised Similarity Learning using Partially Ordered Sets}
  author={Bautista, Miguel A and Sanakoyeu, Artsiom and Ommer, Bj{\"o}rn},
  booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2017}
}