Presentation on theme: "Large-Scale, Real-World Face Recognition in Movie Trailers Alan Wright."— Presentation transcript:
Large-Scale, Real-World Face Recognition in Movie Trailers Alan Wright
Plan of Attack Extract Facial Tracks from videos Extract the features from the facial tracks. Build framework to load and test data. Begin with baseline testing (Sparse, min, mean, etc) Algorithm development…
Dataset 2010 Movie Trailers o 305 videos o Avg. of 625 faces per trailer o Avg. of 23 face trackers per trailer PubFig o 58,787 photos o 200 people o Avg. of 295 photos per person.
Preliminary Testing Ran SRC facial recognition on extracted facial tracks from Date Night. Choose the name that has the most votes in the track. Looking strictly at Steve Carell and Tina Fey tracks…
Results Correctly assessed 15 out of 34 face tracks correctly (44.12% accuracy). 6 out of 18 Tina Fey face tracks (33.33% accuracy). 9 out of 16 Steve Carell face tracks (56.25% accuracy).
Correct Result Track 7- IDed CORRECTLY as: Tina Fey o Alyssa Milano had 2 votes o Drew Barrymore had 4 votes o Jennifer Love Hewitt had 1 votes o Salma Hayek had 3 votes o Tina Fey had 22 votes 32 frames in this track 68.75% identified as Tina Fey
Incorrect Result (3) track 40 - IDed as: Keira Knightley o Adriana Lima had 6 votes o Cindy Crawford had 8 votes o Eliot Spitzer had 1 votes o Eva Mendes had 1 votes o Gillian Anderson had 1 votes o Keira Knightley had 18 votes o Meg Ryan had 3 votes o Minnie Driver had 1 votes o Salma Hayek had 3 votes o Tina Fey had 15 votes
Dark lighting Pose variation Motion blur toward the end 57 frames
Whats next? Continue to look at the coefficient vector. Look into confidence of each choice (SCI and residual error) Apply different facial recognition algorithms. Best way to choose a name for each track? (Mode, strongest confidence, etc.)