UQMG @ TRECVid 2015: Instance Search

The UQMG group submits three runs for instance search at TRECVid 2015 [13]: all of them are automatic runs. Instead of adopting the traditional retrieval approach, e.g., Bag-of-Visual-Word (BoVW), our approach consists of three major steps: video decomposition, feature extraction and indexing. During decomposition, video segmentation is applied and various objects are extracted. Here a visual object is a minimal unit, and a video might consists of thousands of objects. Then we extract the visual feature of the object by using a convolutional neural network (ConvNet), which is a high-dimensional vector outputted by a fully connected layer of the network. Finally, the instance search problem is treated as finding the approximate nearest neighbors (ANN) of a given query in a large set of data points in high-dimensional space. Our best mAP is 0.114.