1 / 15

Large-Scale, Real-World Face Recognition in Movie Trailers

Large-Scale, Real-World Face Recognition in Movie Trailers. Presentation 4 Alan Wright. Dictionary Distribution. Number of training images (capped at 200). Pub Fig Dictionary (200 people). Dictionary Distribution. Number of training images (capped at 200). Pub Fig Dictionary (200 people).

leola
Télécharger la présentation

Large-Scale, Real-World Face Recognition in Movie Trailers

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Large-Scale, Real-World Face Recognition in Movie Trailers Presentation 4 Alan Wright

  2. Dictionary Distribution Number of training images (capped at 200) Pub Fig Dictionary (200 people)

  3. Dictionary Distribution Number of training images (capped at 200) Pub Fig Dictionary (200 people)

  4. Preliminary Testing • GPSR – Date Night 76.47% accuracy (26/34 tracks) • Accuracy increase from 44% last week. • Looked at confidence rather than only mode. • Adjusted tau parameter to 0.01 • Accurate results, but takes ~320 secs per frame (approx 5 minutes). Face Tracks have anywhere from 8 to 80+ frames.

  5. Preliminary TestingCoefficient Vector Test images in dictionary Tina Fey Frames in Track

  6. Preliminary TestingCoefficient Vector • Track is Tina Fey • High confidence and high coefficients • Exactly what we expected

  7. Preliminary TestingCoefficient Vector Test images in dictionary Frames in Track

  8. Preliminary TestingCoefficient Vector • No clear recognition • Most likely due to facial expression, lighting, pose, and small number of frames (only 8).

  9. Preliminary Testing • Other methods tested: homotopy, DALM, LASRC… • Poor accuracy, but faster run times. • How can we speed up GPSR?

  10. GPSR Average Approach Y1 = Ax1Y2 = Ax2 …Yk = Axk Simplifies to:Find a single coefficient for an entire track, rather than a coefficient for each frame of the track.

  11. GPSR Average Approach • Average each frame of the track (very quick). • Pass this average to the face recognition function. • It will still take ~320 secs, but won’t take 320 s * x frames.

  12. GPSR Average Results • Date Night – 76.47% accuracy (just as accurate as the frame by frame method, but faster). • The incorrect matches were on all but 2 of the same face tracks.

  13. GPSR Average Results • Preformed just as well on trailer subset (6 trailers) • George Bush was IDed in all 3 tracks. • Celine Dion was IDed in 4 out of 6 tracks.

  14. GPSR Average Results • Jessica Alba was not IDed well • Leonardo DiCaprio was not IDed well

  15. What’s Next • Finalize Dataset – final face tracks and features. • Larger Dataset – Test on more trailers. • Look at additional baseline method – low rank approximation (65% accuracy – 60 secs a track) • Derive more interesting approximations

More Related