Robotics perception

DensePol — dense-angle polarization dataset

A polarization dataset captured through 180 analyzer orientations per scene, making label quality measurable for learned polarimetric vision.

DensePol — dense-angle polarization dataset

Learned polarimetric vision is only as good as its supervision, yet most RGB–polarization datasets take their labels from four-angle snapshot sensors. DensePol captures each scene through 180 full-resolution analyzer orientations, a 45-fold denser angular budget, and releases the complete angular stacks and fit residuals rather than only derived maps (524 sweeps; 2,018 paired RGB–polarization images).

The dataset makes label quality measurable: relative to the dense reference, angle-of-polarization labels from standard four-angle capture deviate by 13.4°, falling to 2.2° at 90 views. It was acquired with the division-of-time sensor built in the lab — a color camera behind a linear polarizer on a motorized rotation stage (shown).

Publication

P. Sangani, A. Moori, E. Blasch, G. Seetharaman, and H. Aliakbarpour. DensePol: Dense-Angle Polarization Dataset for Learning-Based Polarimetric Vision. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2027 (accepted). arXiv