No abstract

COCO Annotator is an image annotation tool that allows the labelling of images to create training data for object detection and localization. It provides many features, including the ability to label an image segment by drawing, label objects with disconnected visible parts, efficiently store and export annotations in the well-known COCO format as well as importing existing publicly available datasets in COCO format. Once installed, or started with Docker, the interface is web-based and customizable, and provides different tools for creating datasets. The exported annotations can be used for the training of modern deep learning algorithms (Mask R-CNN, YOLO, etc.). Import of the COCO format is supported by several machine learning frameworks. COCO Annotator is developed by Justin Brooks and is supported by a lively community on Github [1].

Figure 1: Full Approach. We use GoogleNet and SIFT on two different data split strategies. Due to very similar images in the extracted frames, the random shuffle k-fold cross validation produces unrealistic high results on multiple metrics. SIFT confirms this similarity. The video-based split shows what we might except in terms of real-world performance.