Open World Object Detection in Satellite Images
This work brings the open-world detection problem for the first time to the field of satellite imagery with the introduction of two novel splits based on the number of object instances to imitate different real-world scenarios. We introduce a Rotation-aware Open-World detection transFormer (ROWFormer) and also adapt CNN-based work Sparse R-CNN, to explore two different directions into this problem. Our work introduces a Rotation-aware Pyramidal Pseudo-Labeling (RaPPL) scheme to effectively capture scale-specific backbone features at oriented box regions to learn and label unknown objects, without explicit supervision, to tackle the challenges specific to aerial imagery. To this end, we try to address these challenges including orientation and clutter with dense objects in satellite images in this work. Comprehensive experiments on our OWOD-S show the benefits of the RaPPL scheme with ROWFormer. Within ROWFormer, the RaPPL scheme provides an absolute gain of 6.4% and 6% in unknown recall over the baseline in our two splits. Further, we also adapt the standard Sparse R-CNN to oriented open-world object detection in satellite images and it provides promising results. Lastly, we also experiment with our ROWFormer in a semi-supervised learning setup. The source code, models, and OWOD-S splits will be made public.
A. S. Gehlot, "Open World Object Detection in Satellite Images", M.S. Thesis, Computer Vision, MBZUAI, Abu Dhabi, UAE, 2023.