Off-the-shelf convolutional neural network features achieve outstanding results in many image retrieval tasks. However, their invariance is pre-defined by the network architecture and training data. Existing image retrieval approaches require fine-tuning or modification of the pre-trained networks to adapt to the variations in the target data. In contrast, our method enhances the invariance of off-the-shelf features by aggregating features extracted from images augmented with learned test-time augmentations. The optimal ensemble of test-time augmentations is learned automatically through reinforcement learning. Our training is time and resources efficient, and learns a diverse test-time augmentations. Experiment results on trademark retrieval (METU trademark dataset) and landmark retrieval (Oxford5k and Paris6k scene datasets) tasks show the learned ensemble of transformations is effective and transferable. We also achieve state-of-the-art MAP@100 results on the METU trademark dataset.
View on arXiv