Unsupervised Visual Representation Learning by Graph-based Consistent Constraints Supplementary Material

1. We show cycle detection results using the proposed unsupervised constraint mining approach on the large-scale ImageNet 2012 dataset as well as three image datasets with generic objects (CIFAR-10), fine-grained objects (CUB200-2011), and scene classes (MIT indoor-67), respectively. 2. We show examples of easy negative image pairs (i.e., image pairs with large Euclidean distance in the feature space) and hard negative sample pairs (i.e., image pairs with large geodesic distance in the k-NN graph). 3. We show additional qualitative results and a detailed quantitative evaluation for the unsupervised feature learning task. 4. We report detailed quantitative results on three image classification datasets for the semi-supervised learning task.