Retrieving Objects, People and Places from a Video Collection: TRECVID'12 Instance Search Task
We participated in 2012's TRECVID instance search task (INS) and wanted to measure how much we can positively impact the performance of a state-of-the-art video retrieval system based on local features and relying solely on content-based retrieval methods. Our agenda consisted in the implementation of some incremental additions to the system that included filtering of local features, custom codebook generation and tailored ranking metrics for the indexed videos. We got three versions of our system tested which iteratively included algorithms that consistently pushed further the performance of the system. Given the terms under which the system had to be implemented - ground-truth was not available-, we used artificially generated datasets to get an idea of how much progress were we making with each additional component. We showed that improvements requiring small computational and human effort can already have positive impacts on the system's performance.