Outlier Detection for Gesture Videos
In the last 3 years, we have collected a dataset of 25000 isolated sign videos belonging to ~650 different Turkish Sign Language gestures. These signs contain repetitions of the same sign from many different users. However, due to the scope of the data, it is hard to detect mistaken samples or classes from these videos one by one. Therefore in this project, we will implement an outlier detection application, that will analyze different characteristics of these videos to find mislabeled videos and correctly label them.
The project involves:
- Extraction of coordinate and appearance based features.
- Implementing a spatio-temporal outlier detection algorithm using temporal analysis, clustering and outlier detection methods.
- Developing detection methods for common scenarios such as:
- User performing a wrong gesture.
- User asking for a repeat after a gesture.
- Detecting recording errors such as timejumps, late starts, early stops.
- Usage of wrong hands.
- Consistency errors.
Project to be guided by Lale Akarun and Ahmet Alp Kındıroğlu.