Better accuracy, lower costs
Problem being addressed
Deep learning has achieved high performance on many object detection tasks like remote sensing detection, pedestrian detection and face detection; while a lot of effort has been made to improve detector’s accuracy, we usually have excellent performance on high quality datasets and relatively poor performance on self-made datasets. At the same time, high accuracy on remote sensing images requires huge labor and time costs of annotation.
A new uncertainty-based active learning algorithm which can select images with more information for annotation and detector can still reach high performance with a fraction of the training images. The method not only analyzes objects’ classification uncertainty to find least confident objects but also considers their regression uncertainty to declare outliers.
Advantages of this solution
The method allows to achieve same-level performance as full supervision with only half images. In the future work semi-supervised module may be added to further reduce annotation cost for detector’s training.
Solution originally applied in these industries
Possible New Application of the Work
Remote sensing images approach is approach is widely used in resource detection and ecological research, the areas that are often underfunded, so a cheaper way to obtain accurate results would be very beneficial in this field.
Source URL: #############