This paper provides a unified perspective of learning with pair similarity optimization, observing that existing loss functions such as triplet loss and softmax cross-entropy loss "embed s_n and s_p into similarity pairs and seek to reduce (s_n - s_p)." The authors propose Circle Loss, which "re-weights each similarity to highlight the less-optimized similarity scores" and yields a circular decision boundary in the similarity space. The approach achieves "a more flexible optimization approach towards a more definite convergence target."

Deep feature learning aims to learn discriminative feature representations for tasks like face recognition, person re-identification, and fine-grained image retrieval. Two dominant paradigms exist: class-level labels used in softmax-based classification, and pair-wise labels used in metric learning approaches like contrastive loss and triplet loss.

Both paradigms ultimately optimize similarity pairs (s_n, s_p), but they treat each pair with equal importance during optimization. Specifically, the gradient contributions are "symmetric for s_n and s_p," meaning a well-optimized similarity score receives the same gradient magnitude as a poorly-optimized one. This leads to less efficient convergence and ambiguous convergence status.

Triplet loss constructs triplets (anchor, positive, negative) and enforces "s_p - s_n > m" where m is a fixed margin. N-pair loss extends this to multiple negatives. Softmax-based methods with angular margins (e.g., ArcFace, CosFace) add penalties to the cosine similarity between features and class centers. Despite their different formulations, all methods share the same underlying optimization of similarity pairs.

The authors first establish a unified loss function for both classification and metric learning: L_uni = log[1 + sum exp(gamma * (s_n - s_p))]. This formulation reveals that all existing approaches optimize "the same similarity pair (s_n, s_p) with the decision boundary s_n - s_p = 0." The key insight is that this linear decision boundary treats all similarity scores uniformly, regardless of their current optimization status.

Circle Loss introduces self-paced weighting factors alpha_n and alpha_p that are "determined by the current optimization status of each similarity score." Specifically, alpha_n = [s_n - O_n]+ and alpha_p = [O_p - s_p]+, where O_n and O_p are the optimal values. This means a similarity score far from its optimum receives larger gradient weight, while a well-optimized score receives diminished gradient. The resulting decision boundary in the (s_n, s_p) space is circular rather than linear, providing "a more definite convergence target" at the point (O_n, O_p).

Circle Loss generalizes existing approaches: when alpha_n = alpha_p = 1, it reduces to the unified loss equivalent to triplet loss. The scale factor gamma and margin m control the "radius and center of the circular decision boundary," offering a flexible framework that subsumes prior methods as special cases.

Experiments span three tasks: face recognition, person re-identification, and fine-grained image retrieval. On face recognition, Circle Loss achieves competitive results on MegaFace with 98.50% rank-1 accuracy. On person re-ID using the Market-1501 dataset, Circle Loss reaches 96.1% rank-1 and 87.4% mAP, outperforming triplet loss and softmax baselines. On SOP (Stanford Online Products) for fine-grained retrieval, Circle Loss achieves 78.3% recall@1, surpassing prior state-of-the-art methods. Convergence analysis shows Circle Loss reaches the target accuracy faster than triplet loss and softmax.

Circle Loss provides a unified perspective on pair similarity optimization, revealing that classification and metric learning share the same underlying structure. By introducing self-paced weighting that emphasizes poorly-optimized similarity scores, the method achieves circular decision boundaries with more flexible optimization and definite convergence targets. Extensive experiments on face recognition, person re-ID, and fine-grained retrieval validate its effectiveness and generality.