Improvements of Face Detection and Recognition Combination of existing methods Bc. Vladimír Míč FI MU 2Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 OutlineOutline Basic terms Face detection and recognition Quality evaluation Avaliable software Face detection improvement Aggregated descriptors Face recognition improvement 3Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Face detection problem Goal is to put an ellipse on the place where a face is Recall: how many faces out of real faces were detected? (100 %) Precision: how many faces out of detected ones represent real faces? (66,7 %) 4Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Face recognition problem In general, the ability to answer the question “is it the same person in these two images?” In practice, sorting faces according to the similarity with respect to a query Similarity expressed via distance function Example: 21 photos of person “00003”, 10-NN query Recall = 4 / 21 = 19 %, Precision = 40 % Distance Person id 0 00003 8932 00003 9145 00003 9167 00003 9277 00750 9281 00765 9282 00972 9283 00695 9285 00772 9286 00750 5Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Available software OpenCV Opensource, MPEG7 descriptors, metric properties Can make descriptor from an arbitrary picture Luxand, Neurotechnology (Verilook) Commercial software, own descriptors Recognition only of faces detected by its own 6Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Detection improvement Based on compliance of pieces of software Compliance means sufficient overlay of detected areas Recommended variant: compliance of at least two out of all pieces of software Precision nearly 100 % Aggregated descriptor Holds several descriptors of the same face 7Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Aggregated face descriptor – our variant Encapsulated objects MPEG7 descriptor Luxand descriptor Verilook descriptor MPEG7 descriptor is always present Can be added via crop made according to the Luxand or Verilook descriptor 8Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 1 1260 small faces, low quality 2 66 big faces, high quality Face detection results Name Recall1 Precision1 Recall2 Precision2 Open CV (OCV) 55 % 89 % 92 % 86 % Luxand 63 % 83 % 95 % 94 % Neurotechnology (Verilook) 73 % 84 % 100 % 96 % Aggregated extractor 62 % 98 % 97 % 100 % 9Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Recognition improvement Distance between Aggregated descriptors Combination of several distances between encapsulated objects (partial distances) Some partial distances may be missing Normalization of each partial distance by a precision Use minimal value of normalized partial distances 10Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Distance normalization Normalization function: norm(d) = 1 – p Illustration: distance 195 is normalized to 1 – 0,6198 = 0,3802 Precision is measured Need of training sample data 11Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Training data problem Cons: Normalization strongly depends on a dataset It's suitable to provide own training sample User must identify faces (abstract id, name, … ) Pros: High precision of face detection may be used Face id in file path for images with one face Training data may be very small to provide good results 12Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Real testing sample – 753 small images 13Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Results with small training datasets FERET: 998 people, 6 379 faces (red: 9 people, 224 faces) 14Improvements of Face Detection and RecognitionBc. Vladimír Míč, jaro 2014 Performance boost Index build on Aggregated descriptors using MPEG7 distance function (needs metric properties) Candidates selection (logarithmic complexity) Overrank according the aggregated distance function Thank you for your attention