Robust and Decomposable Average Precision for Image Retrieval
This work addresses a specific challenge in image retrieval for researchers and practitioners by improving end-to-end training with AP, though it is incremental as it builds on existing AP approximation methods.
The paper tackles the problem of training deep neural networks with average precision (AP) for image retrieval by addressing its non-differentiability and non-decomposability, resulting in a method that outperforms recent AP approximations and achieves state-of-the-art results on three datasets.
In image retrieval, standard evaluation metrics rely on score ranking, e.g. average precision (AP). In this paper, we introduce a method for robust and decomposable average precision (ROADMAP) addressing two major challenges for end-to-end training of deep neural networks with AP: non-differentiability and non-decomposability. Firstly, we propose a new differentiable approximation of the rank function, which provides an upper bound of the AP loss and ensures robust training. Secondly, we design a simple yet effective loss function to reduce the decomposability gap between the AP in the whole training set and its averaged batch approximation, for which we provide theoretical guarantees. Extensive experiments conducted on three image retrieval datasets show that ROADMAP outperforms several recent AP approximation methods and highlight the importance of our two contributions. Finally, using ROADMAP for training deep models yields very good performances, outperforming state-of-the-art results on the three datasets.