Camera clustering for scalable stream-based active distillation

Dani Manjah, Davide Cacciarelli, Christophe De Vleeschouwer, Benoît Macq

Abstract: Highlights•We introduce a scalable framework for training and deploying lightweight DNNs in multi-camera systems.•It collects images from each camera, labels them using a large general-purpose model, and clusters the data based on camera similarity.•We show that selecting images where a model predicts with high confidence reduces querying cost, while increasing model accuracy post-training.•We propose camera clustering distance based on the fact that a model fine-tuned for a camera transfers to another camera with similar features.•Clustering reduces the number of models to train and enhances accuracy by striking a trade-off between one model per camera and one for all cameras.

External IDs:doi:10.1016/j.eswa.2025.128408