DCM2Net: an improved face recognition model for panoramic stereoscopic videos

Area of Science:

• Computer Vision
• Artificial Intelligence
• Multimedia Technology

Background:

• Panoramic stereo video offers immersive experiences but presents significant facial deformation challenges.
• Accurate face recognition in such videos is crucial for applications but is hindered by image distortions.

Purpose of the Study:

• To develop an effective face recognition model specifically for panoramic stereo videos.
• To address the challenges posed by facial deformation in immersive video content.

Main Methods:

• Proposes the DCM2Net (Deformable Convolution MobileFaceNet) model.
• Integrates inter-channel feature information during fusion and redistribution for enhanced feature extraction.
• Developed a live panoramic stereo video system incorporating the DCM2Net model for real-time recognition.

Main Results:

• The DCM2Net model demonstrates superior performance on benchmark and panoramic datasets.
• Experimental results validate the model's effectiveness in recognizing faces within deformed panoramic stereo video.

Conclusions:

• DCM2Net offers a robust solution for face recognition in panoramic stereo video.
• The model's ability to handle feature deformation significantly improves recognition accuracy in immersive media.