IWBF 2014 2nd International Workshop on Biometrics and Forensics 27‐28th March, Valletta, Malta TOWARDS PRACTICAL SPACE‐VARIANT BASED FACE RECOGNITION AND AUTHENTICATION Enrico Grosso, Andrea Lagorio, Luca Pulina, and Massimo Tistarelli Computer Vision Laboratory - University of Sassari, ITALY CONTEXT AND MOTIVATION Combining the extraction of salient points and space‐variant descriptors, a novel and efficient method for face recognition and authentication can be implemented. This method is called BSV. Experimental results demonstrate that this approach is computationally feasible and guarantees a significant increase in accuracy with respect to the original SIFT approach. THE BOOSTED SPACE VARIANT FRAMEWORK BSV(img1, img2) 1 kp1 = SIFTDETECT(img1) 2 kp2 = SIFTDETECT(img2) 3 desc1 = SIFTEXTRACT(kp1) 4 desc2 = SIFTEXTRACT(kp2) 5 skp1 = NIL 6 skp2 = NIL 7 SIFTMATCHER(desc1, desc2, skp1, skp2) 8 score = ‐1 9 for i = 0 to SIZE(skp1) 10 currScore = LPCORR(skp1[i], skp2[i]) 11 if currScore > score then 12 score = currScore 13 return score Example of a single space variant filter centered on a generic point RESULTS FACE RECOGNITION (FERET DATABASE) Accuracy BSV GSIFT SIFT 82% 78% 72% FACE AUTHENTICATION (BANCA DATABASE) MC UD UA P MC UD UA P BSV 7.12% 23.22% 19.43% 23.31% BSV 3.73% 24.64% 19.00% 21.45% GSIFT 7.02% 12.43% 18.32% 18.09% GSIFT 2.00% 12.82% 17.18% 15.55% SIFT 12.12% 39.44% 28.37% 33.72% SIFT 7.82% 42.18% 29.55% 38.91% WER (1) WER (0.1) MC UD UA P BSV 10.27% 21.36% 19.00% 25.54% GSIFT 11.00% 10.36% 18.82% 20.45% SIFT 15.18% 34.82% 34.45% 29.09% WER (10)