CanonicalPhys fixes head-pose collapse in deep remote heart-rate sensing

The paper's contribution isn't just better accuracy under rotation: it's an argument that the field has been solving the wrong problem. Treating pose variation as a data distribution issue (more augmentation, more diverse training sets) misses that head movement structurally violates the optical physics rPPG models are built on, so no amount of additional training data fixes the underlying mismatch.

The physics-prior thread running through this week's coverage is hard to ignore. PEARL (the physics-enhanced RL paper from the same day) makes a structurally similar argument: that injecting domain-specific physical constraints into a learning system beats brute-force data scaling when the problem has known geometric or dynamic structure. CanonicalPhys applies the same logic to a biometric sensing pipeline, using anatomical geometry to regularize what the model sees rather than hoping learned features generalize across poses. Neither paper is claiming physics replaces data entirely; both argue that the right prior narrows the hypothesis space in ways that matter for real deployment.

The meaningful test is whether CanonicalPhys holds up on datasets with combined pose and lighting variation, not just pose alone. If the canonical-space normalization degrades under low-light or partial occlusion conditions that appear in the MMPD benchmark's harder splits, the geometric prior may be solving one failure mode while exposing another.