Stochastic Deep Supervision for Gallery-Consistent Early-Exit Face Recognition
Early-exit face matching retains over 96% of full-depth performance at half the compute, using the existing reference gallery.
The authors’ SD-ViT framework trains a randomly selected intermediate block alongside the final layer on each iteration. A shared embedding head and classifier align their identity representations, allowing early-exit probes to match directly against full-depth gallery embeddings. Full-depth performance stays within 0.4 IJB-C points of the conventionally trained backbone. A verification threshold calibrated at full depth remains usable from quarter depth onward.
HF Daily Papers' note
The authors’ SD-ViT framework trains a randomly selected intermediate block alongside the final layer on each iteration. A shared embedding head and classifier align their identity representations, allowing early-exit probes to match directly against full-depth gallery embeddings. Full-depth performance stays within 0.4 IJB-C points of the conventionally trained backbone. A verification threshold calibrated at full depth remains usable from quarter depth onward.
HF Daily Papers' note
score 4