Learning test-time augmentation for content-based image retrieval

Off-the-shelf convolutional neural network features achieve outstanding results in many image retrieval tasks. However, their invariance to target data is pre-defined by the architecture and training data. Existing approaches require fine-tuning or modification of pre-trained networks adapt variations unique In contrast, our method enhances off-the-shelf aggregating extracted from images augmented at test-time, with augmentations guided a policy learned through reinforcement learning. The assigns different magnitudes weights selected transformations, which are list transformations. Policies evaluated using metric learning protocol learn optimal policy. model converges quickly cost each iteration minimal as we propose an off-line caching technique greatly reduce computational extracting images. Experimental on large trademark (METU dataset) landmark (ROxford5k RParis6k scene datasets) tasks show that ensemble transformations highly effective for improving performance, practical, transferable.