SD-ViT trains every ViT block as a usable face-recognition exit
Original title: Stochastic Deep Supervision for Gallery-Consistent Early-Exit Face Recognition
Early-exit probes keep over 96% of full-depth accuracy at half the compute without rebuilding the gallery.
SD-ViT is a training framework that makes every Vision Transformer block produce a discriminative face-recognition embedding for budget-selected early exits. A randomly chosen intermediate block shares the final embedding head and a weighted margin-penalty softmax loss, so early probes can be matched to full-depth gallery embeddings. On established benchmarks, full-depth SD-ViT stays within 0.4 IJB-C points of conventional training. A threshold calibrated at full depth remains usable from quarter depth onward.
Unconfirmed: Benchmark margins are reported by the paper and not independently confirmed here.
- Early exits recover over 96% of their own full-depth performance at 50% compute.
- No gallery recomputation or adaptation layers are required.
- One full-depth verification threshold stays usable from quarter depth.
- Intermediate blocks are trained with a shared classifier and embedding head.
From the original article
1 min read at huggingface.co
Vision Transformers (ViTs) perform strongly in face recognition (FR), yet their fixed-depth inference limits deployment in resource-constrained settings. We propose SD-ViT, a training framework that gives a usable exit at every block depth by making every depth produce discriminative embeddings, for budget-selected fixed-depth exits.
By Guray Ozgur, Tahar Chettaoui, Eduarda Caldeira · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.