A High Precision and High Recall Face Detector for Equi-Rectangular Images

Ankit Dhiman, Praveen Agrawal

Abstract: 360^{\circ }$$ cameras have recently become quite popular as it captures the entire surrounding environment in one frame. The captured image, however, suffers from spherical distortions because of the fish eye lenses used in these cameras. Equi-rectangular (ER) projections are the most popular and convenient 2D projections when it comes to dealing with $$360^{\circ }$$ images. However, these projections suffer from heavy deformations in the pole regions which distorts the parts of the image falling near pole regions, thus increasing the difficulty of face-detection in these regions. In this paper, we address face-detection in $$360^{\circ }$$ images in their ER projection. We discuss how a detector can be trained efficiently on ER images. We also introduce a deep learning pipeline which has high precision and high recall despite the presence of these spherical distortions.