From Embeddings to Ensembles: A Controlled Comparison of ArcFace and Global CNN Features

This study offers a controlled comparison of facerecognition pipelines on Indonesian Muslim Face Dataset under a single, fixed protocol designed to minimize confounds. Faces are detected and aligned with Multi-task Cascaded Convolutional Network, data are split once using a stratified 80:20 partition. The training split undergoes class balancing and moderate occlusion exposure, while the test split remains untouched and is evaluated with single-crop inference. We examine two representation families: identity-preserving embeddings from ArcFace (ResNet-100, buffalo_l) and frozen global descriptors from EfficientNet-B0 and DenseNet-121. Each representation is paired with four classical classifiers, Support Vector Classifier, K-Nearest Neighbors, Random Forest, and Decision Tree, and assessed using macro-averaged accuracy, precision, recall, and F1 score. Results show that ArcFace produces compact, well-separated identity clusters, yielding near ceiling performance across classifiers, whereas frozen global descriptors are weaker and markedly more sensitive to the choice of decision rule (ensembles tend to suit EfficientNet-B0, while a margin-based SVC aligns better with DenseNet-121). Confusion matrix analyses corroborate these trends, revealing sparse residual confusions for ArcFace and broader, asymmetric error patterns for the global descriptors, especially under partial occlusion. The study offers a reproducible baseline and practical guidance for small-tomedium, imbalanced settings: identity-preserving embeddings coupled with a simple linear margin or KNN provide a dependable default, where only generic CNNs are available, tree ensembles and margin-based methods should be chosen to match descriptor geometry.