Physically-admissible polarimetric data augmentation for road-scene analysis

Polarimetric imaging, along with deep learning, has shown improved performances on different tasks including scene analysis . However, its robustness may be questioned because of the small size training datasets. Though issue could solved by data augmentation, polarization modalities are subject to physical feasibility constraints unaddressed classical augmentation techniques. To address this issue, we propose use CycleGAN, an image translation technique based generative models that solely relies unpaired data, transfer large labeled road datasets polarimetric domain. We design several auxiliary loss terms that, alongside CycleGAN losses, deal images. The efficiency solution is demonstrated object detection where generated realistic images allow improve cars and pedestrian up 9%. resulting constrained publicly released, allowing anyone generate their own • first framework for generating polarization-encoded. a handling in show these help improving object-detection task.