Authors:
Alexandru Ghita
and
Radu Tudor Ionescu
Affiliation:
Department of Computer Science, University of Bucharest, 14 Academiei, Bucharest, Romania
Keyword(s):
Contrastive Learning, Contrastive Loss, Learnable Class Anchors, Content-Based Image Retrieval, Object Retrieval.
Abstract:
Loss functions play a major role in influencing the effectiveness of neural networks in content-based image retrieval (CBIR). Existing loss functions can be categorized into metric learning and statistical learning. Metric learning often lacks efficiency due to pair mining, while statistical learning does not yield compact features. To this end, we introduce a novel repeller-attractor loss based on metric learning, which directly optimizes the L2 metric, without pair generation. Our novel loss comprises three terms: one to ensure features are attracted to class anchors, one that enforces anchor separability, and one that prevents anchor collapse. We evaluate our objective, applied to both convolutional and transformer architectures, on CIFAR-100, Food-101, SVHN, and ImageNet-200, showing that it outperforms existing functions in CBIR.