Classification of human actions in video 

Practical session INRIA CVML Summer School, Grenoble, July 26-30, 2010

Ivan Laptev, INRIA/ENS, Paris, France

Part I: Interest points and visual vocabulary

- Start Matlab in cvml10-actions folder
- Follow instrictions in script01.m
- Extract Space-Time Interest Points (STIPs) on two example sequences
- Display results:



- K-means clustering of feature descriptors from the first sample sequence
- Label assignment of features from the second sequence

Part II: Action recognition

- Start Matlab in cvml10-actions folder
- Follow instrictions in script02.m
- Display pre-computed STIP features on a sample movie clip

- Compute BOF representation of videos from pre-quantized STIP features
- Compare performance of Nearest Neighbour and SVM classifiers for the problem of discriminating two action classes: "Hug Person" and "Sit Down" in the challenging setup of movie clips in Hollywood-2 dataset  
- Compare performance to a random classifier

Precision-Recall for "Hug Person"

Precision-Recall for "Sit Down"