From Gradient Analysis to Norm Control: Rethinking Triplet Loss for Gait Recognition

Gait recognition has attracted increasing attentionin both academia and industry as a non-intrusive human recog-nition technology from a distance without requiring cooperation.Triplet loss, which enforces relative distance constraints, is afundamental component in gait recognition. Recently, severalgait-specific triplet losses have been introduced to gait recogni-tion. However, they only focus on sample selection and weightingto enhance constraints without exploring the gradient propertiesof Cosine/Euclidean metric, which fundamentally influence themodel training efficiency and feature discriminability. In thispaper, we theoretically analyze triplet loss gradients combinedwith weight decay and identify inherent limitations due to inad-equate norm-control: Cosine metric triplet loss (L_cos) exhibitsexcessive gradients resulting from small feature norms, whileEuclidean metric triplet loss (L_euc) suffers from a small margin-to-norm ratio due to large feature norms. To address these issues,we propose two norm-control approaches to constrain the featurenorm in a stable range: 1) Norm-Variance-Regularized Collab-oration. 2) Norm-Based Regularization. Extensive experimentsshow that our methods outperform state-of-the-art results underboth Cosine and Euclidean evaluation metrics on three in-the-wild datasets: Gait3D, GREW, and BUAA-Duke-Gait.