SLIDE 1
Action Recognition with Improved Trajectories
Heng Wang and Cordelia Schmid LEAR, INRIA, France
Introduction
- Problem
○ Action recognition - Classify a set of frames into a motion. What is he doing? [UCF Sport dataset]
Action Recognition with Improved Trajectories Heng Wang and Cordelia - - PowerPoint PPT Presentation
Action Recognition with Improved Trajectories Heng Wang and Cordelia - - PowerPoint PPT Presentation
Action Recognition with Improved Trajectories Heng Wang and Cordelia Schmid LEAR, INRIA, France Introduction Problem Action recognition - Classify a set of frames into a motion. What is he doing? [UCF Sport dataset] Introduction
SLIDE 2
SLIDE 3
Introduction
- Difficulties
○ Motion blur ○ Background trajectories [UCF Sport dataset]
SLIDE 4
Introduction
- How do we improve noisy trajectories?
○ Estimate camera motion ○ Human detector [Hollywood2]
SLIDE 5
Introduction
[Hollywood2]
SLIDE 6
Background
- Motion-based Descriptors
○ HOF ○ MBH ○ 3D SIFT ○ Extended SURF ○ HOG3D [Chaudhry et. al, OpenCV]
SLIDE 7
Background
- Approach
○ Approximate camera ■ SURF ■ Good Features to Track [Opencv documentation]
SLIDE 8
Background
- Approach
○ WarpFlow ■ warp optical flow ○ RmTrack ■ remove background [Hollywood2]
SLIDE 9
Experiment
- Datasets
○ UCF50 ■ Youtube ■ Semi-cluttered ○ HMDB51 ■ Most challenging ■ Varies in camera, quality [UCF101]
SLIDE 10
Experiment
- Visual Comparison
○ Baseline - Dense Trajectories ○ Camera estimation + human mask
- Demo
[Hollywood2]
SLIDE 11
Experiment
- How do descriptors do?
○ HOF ○ HOG ○ MBH [Hollywood2]
SLIDE 12
Experiment
HOF HOG MBH Baseline Dense Trajectories Stab
SLIDE 13
Experiment
HOF HOG MBH Baseline Dense Trajectories Stab
SLIDE 14
Experiment
- Failure cases
○ Motion blur ○ Illumination changes ○ Lots of humans [HMDB51]
SLIDE 15
Experiment
- Failure cases
○ Motion blur ○ Illumination changes ○ Lots of humans [HMDB51] Why? Recall how we estimate camera motion - SURF
SLIDE 16