90 likes | 224 Views
Action Recognition. Karthik Prabhakar UCF REU 2008, Week 8 Report July 11, 2008. Experiment Methods: UCF Sports Actions. Final (GOOD) Results 1: One-Against-All Linear SVM
E N D
Action Recognition Karthik Prabhakar UCF REU 2008, Week 8 Report July 11, 2008
Experiment Methods: UCF Sports Actions • Final (GOOD) Results 1: • One-Against-All Linear SVM • Parameters C and ε were found through exhaustive search on intervals C = [2-15,2+15] and ε = [2-5,2+5]. • Experiments were carried out using a 2-fold and 5-fold cross-validation. • Final (GOOD) Results 2: • One-Against-One Linear SVM • (for now, C and ε were chosen from the best run of above) • Experiments were carried out using a 2-fold and 5-fold cross-validation. • Intermediate Results: • A similar framework to above was used for comparison purposes.
Intermediate Results: Linear Binning • Experiments were carried out WITHOUT the use of log-polar3D binning. • That is, only ‘Stage 1’ of the motion descriptor was used. (‘Stage 2’ is the log-polar3D binning approach…our contribution in this paper): Motion Descriptor: Stage 1
Secondary Results: Magnitude Angle • Initial experiments on using optical flow’s magnitude and angle features. Motion Descriptor for Magnitude-Angle: Stage 1
Magnitude-Angle : Intermediate Results • Log-polar3D binning approach vs. Linear binning under the Magnitude-Angle framework.