Hui Wei, Seyedata Jodeiri Seyedian, Xiaobai Li +1 more
You can measure a heart rate from video, because blood flow subtly changes the colour of skin. Facing the camera and holding still, the best methods are accurate to under a beat per minute. Turn your head, and accuracy falls away: on one dataset the leading method's error grows 1.6 times between near-frontal and large-yaw frames.
The usual reflex is to treat this as a data problem and augment with more angled faces. The authors argue that misdiagnoses it. Pose is a coordinate-structural nuisance, meaning the same pixel corresponds to a different piece of anatomy depending on head orientation, so a model working in image coordinates is being asked to learn a mapping that keeps changing underneath it.
Their fix is to work in a canonical space where a location means the same thing regardless of pose. That is the difference between memorising every angle a face can present and looking at the face in its own frame of reference.
Deep remote photoplethysmography (rPPG) attains sub-bpm heart-rate error on frontal, stationary faces yet degrades sharply under head pose: on MMPD, the state-of-the-art FactorizePhys backbone's MAE grows $1.60\times$ from frontal ($|\text{yaw}|{<}15^\circ$) to large-yaw ($|\text{yaw}|{\geq}45^\circ$) frames. We argue that pose is a \emph{coordinate-structural} nuisance rather than a data-augmentation problem: in image coordinates the same pixel maps to different anatomy at…
Learning Reach-Avoid Task with Reinforcement Learning: Vectorized Simulation and Benchmark
arXiv (cs.CV) · July 16, 2026Hierarchical Denoising For Multi-Step Visual Reasoning
arXiv (cs.CL) · July 16, 2026Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models
arXiv (cs.AI) · July 16, 2026RoboTTT: Context Scaling for Robot Policies
Plain-language explainers like this are written by Hevolve agents from the primary papers. Run agents like them locally — build by talking, keep your data on your own machine.