A novel tool for ground truth data generation for video-based object classification

The automatic detection of objects of interest from video sequences is a task of great importance for computer vision applications. Mostly, the detection of these objects is performed using supervised classification techniques, requiring a set of labeled training samples, known as ground truth data. These training samples are commonly labeled by a manual process using rectangular regions (bounding boxes), which is a very tedious task. In this sense, some tools to support the video labeling have been developed, all of them aiming to make more automatic the process using video processing techniques. Nonetheless, these tools lack of a modular implementation, so the users can not include their own techniques to replace or complement the ones used inside the tool. Moreover, none of these tools include a feature estimation stage inside, so it is required to calculate the respective characteristics aside, which might not be a straightforward task. Regarding this, we developed a tool named GTGenerator to support the ground truth data annotation from video streams. GTgenerator is an open source tool, in which the users can easily customize or include their own techniques to support the labeling process. Besides, GTgenerator allows the user to perform the feature estimation inside the same tool, which avoids further processes, making it an ideal tool for object detection. We employed the proposed tool for labeling and extracting features for a fish classification task. The obtained results show that the proposed GTGenerator supports the labeling process, and thus, the classification, attaining a high performance.