RESEARCH — ONE LINE, IN MOTION
estimating stress from facial video and physiological signals — reproduced faithfully, evaluated honestly.
Reproducing and honestly re-evaluating multimodal stress-task classification (CCT-LSTM with EDA augmentation) on UBFC-Phys — the journal version is in preparation. Vision-language directions are planned next.
The first public stop on this line — contact signals join the video.
Reproduces CCT-LSTM for stress-task classification on UBFC-Phys and augments it with wrist-worn EDA under a subject-level protocol. Preliminary by design — the journal version extends and re-evaluates it.
Why contact signals stay: rPPG degrades exactly when stress happens.
Audits how remote photoplethysmography behaves when the subject is actually stressed — motion, expression, and autonomic load degrade the signal precisely in the moments that matter. The audit outlived its venue: it now motivates the journal version.
New architecture, strict protocol, full ablations.
The consolidated account: a faithful reproduction of the baseline, a new architecture, a leakage-safe re-evaluation under subject-level splits and multiple seeds, and an honest diagnostic of what fine-grained stress levels the data can and cannot support — written to be reproduced, fully open on release.
The headline number lands with the journal — until then, the protocol is the promise.