Reducing the Pain: A Novel Tool for Efficient Ground-Truth Labelling in Images

Machine Learning solves more image processing problems every year, yet it is still reliant on painstaking manual ground truth labelling. Segmentation labels require higher accuracy and more clicks than bounding boxes or classification labels. To accelerate the labelling task, a More Efficient Labelling Tool (MELT) has been developed which incorporates features from existing tools and adds some novel ones. The new features are automatic zoom to existing bounding boxes and tracking of arbitrarily shaped objects. Zooming to bounding boxes makes it easy to upgrade bounding box labels to segmentation masks, or to label parts of an object, such as lights on a vehicle. Tracking is available in other tools for rectangular objects such as bounding boxes, but many objects including vehicle lights are not rectangular. The user is given the freedom to create labels with a brush, polygon or superpixel, with customisable label names and colours. Using MELT, a dataset of over 800 images has been prepared with image segmentation labels for vehicle head lights and tail lights. Labels are provided for download as mask files. As there is currently no comparable dataset available, it is hoped that this will become a benchmark for researchers working on detecting and tracking vehicle lights.