Code and Demos

Robust, Real-time Object Recognition

I wrote a robust, real-time object recognition system that uses linear SVM classifiers to do both object recognition and pose classification. The system can also incrementally update the classifiers in an online fashion using warm start, without having to retrain the classifiers from scratch each time you add new data. The source code is adapted from Liblinear, a popular SVM library, and is available here. Most of my modifications are in obj_recog_demo.cpp, obj_recog_demo.h and linear.cpp.

Below is a video demonstrating this system applied to classifying common household objects, including the online learning of a new object.

The system has also been deployed as part of an interactive LEGO playing scenario that was exhibited in the Intel booth during the 2011 International Consumer Electronics Show (CES). In addition to object recognition, we also included pose estimation in this system. Taken together, this allows us to project the correct animations on the objects, including fire from the dragon's mouth, the road in front of the house, water from the fire truck, and the arrow for laying tracks in front of the train. Below is a video demonstration recorded from the event: