params.json

{
  "name": "CliqueCnn",
  "tagline": "Code for our paper \"CliqueCNN: Deep Unsupervised Exemplar Learning\" https://arxiv.org/abs/1608.08792",
  "body": "# Project page of CliqueCnn\r\n\r\nExemplar learning is a powerful paradigm for discovering visual similarities in\r\nan unsupervised manner. In this context, however, the recent breakthrough in\r\ndeep learning could not yet unfold its full potential. With only a single positive\r\nsample, a great imbalance between one positive and many negatives, and unreliable\r\nrelationships between most samples, training of Convolutional Neural networks is\r\nimpaired. Given weak estimates of local distance we propose a single optimization\r\nproblem to extract batches of samples with mutually consistent relations. Conflict-\r\ning relations are distributed over different batches and similar samples are grouped\r\ninto compact cliques. Learning exemplar similarities is framed as a sequence of\r\nclique categorization tasks. The CNN then consolidates transitivity relations within\r\nand between cliques and learns a single representation for all samples without\r\nthe need for labels. The proposed unsupervised approach has shown competitive\r\nperformance on detailed posture analysis and object classification.\r\n---\r\nAll our models fro Olympic Sports dataset can be downloaded from [here](https://hcicloud.iwr.uni-heidelberg.de/index.php/s/kRp6b454Dd0wnts)\r\n\r\nModels [deploy.prototxt](olympic_sports_retrieval/models/deploy.prototxt)  \r\n\r\nEvaluation script [calculate_roc_auc.py](olympic_sports_retrieval/calculate_roc_auc.py)",
  "note": "Don't delete this file! It's used internally to help with page regeneration."
}