Learning from Less: Pose-Balanced Sampling for High-Performance Face Recognition

Large-scale face recognition (FR) datasets are strongly skewed towards frontal images, which limits the ability of models to learn pose-robust representations and leaves frontal-profile verification as a persistent source of error. This paper proposes a pose-balanced sampling strategy that builds a compact training subset by conditioning selection on the per-identity head-pose distribution rather than on sample difficulty or redundancy. Yaw angles are estimated with 6DRepNet, and an equal number of frontal and profile images are retained for every identity. Applied to MS1Mv3, the strategy yields MS1Mv3-FP, a subset of 547,202 images (10.7% of the original) covering 78.5% of the identities. An SE-ResNet100 trained from scratch on MS1Mv3-FP using the ArcFace loss reaches up to 98.73%, 93.23%, and 94.74% verification accuracy on CFP-FP, CPLFW, and VGG2-FP (mean over three sampled subsets: 98.63%, 93.13%, and 94.50%), within 0.30, 0.70, and 0.86 percentage points (pp) of the same network trained on the complete dataset, while reducing training time from more than 100 h to approximately 11 h on a single GPU. Equal-size subsets built without pose balancing lose 0.8–3.6 pp on average, indicating that the pose composition of the subset, and not only its size, accounts for this result. Frontal–frontal verification accuracy is largely preserved, whereas age-gap verification degrades, delimiting the scope of the strategy. The resulting reduction in training time and data could support the development of pose-robust FR models for camera-based sensing systems operating under resource constraints.