Sparse Point Annotations and Iterative Active Learning for Vehicle Detection

This research investigates the application of iterative point-based sparse annotations for semantic segmentation of cars in remote-sensing imagery, to mitigate the challenges associated with laborious and expensive data labeling processes. Car labeling considered the selection of point shapefiles in the Geographic Information System with a value of 1 for the specific class of car, 0 to background and a value of -1 outside the intended target. The semantic segmentation model was the U-Net architecture, with an Efficient-net-B7 backbone and a modified cross-entropy loss function. The experimental evaluation uses the BSB Vehicle Dataset, encompassing two classes (background and vehicles). The results showcase promising improvements, particularly in error-prone classes, as more samples are iteratively added during training. This approach presents a viable and time-efficient alternative for dataset creation, leveraging sparse annotations that are incrementally enhanced. Our pipeline included five iterative rounds in which the IoU increased more than 30% from the first round to the fifth, achieving 60% IoU with 0.059% of the total annotated data. Showing a very good performance with less than 0.1% of pixels annotated. This research advances the field by proposing a rapid and cost-effective method for generating high-quality datasets in remote sensing.