This is the PyTorch implementation of the improved version of our paper Domain Generalized Object Detection for Remote Sensing Images
Building roof type detection from remotely sensed images is a crucial task for many remote sensing applications, including urban planning and disaster management. In recent years, deep learning-based object detection approaches have demonstrated outstanding performance in this field. However, most of these approaches assume that the training and testing data are sampled from the same distribution. When there are differences between the distributions of training and test data, known as domain shift, the performance significantly degrades. In this paper, we proposed a domain generalization method to address domain shift at the instance and image level for roof type detection from remote sensing images. We incorporated consistency regularization to enforce uniformity between these two levels of domain generalization. Furthermore, we evaluated our proposed method with IEEE Data Fusion Contest 2023 dataset. The proposed approach is the first of its kind in terms of domain generalization for remote sensing object detection.
In the context of the IEEE Data Fusion Contest 2023, a large-scale remote sensing image dataset has been provided for the detection of building roof types. The images were collected from the SuperView-1 and Gaofen-2 satellites, with spatial resolutions of 0.5m and 0.8m, respectively. The images in the dataset have a spatial size of 512×512 and were gathered from seventeen different cities across six continents and are labeled based on twelve distinct categories of roof types. We selected three cities in Asia as the validation set, four cities in South America as the test set, and ten cities from the remaining four continents as the training set. Each city represents a different domain. The roof type distribution in the dataset is highly imbalanced. In order to avoid dealing with the imbalanced class problem, we utilized this dataset for building detection, treating all these twelve distinct categories as a single class representing buildings.
| image | instance | consistency | mAP | mAP@50 | |
|---|---|---|---|---|---|
| Faster R-CNN | 0.106 | 0.223 | |||
| DGFRCNN-RS | ✓ | 0.108 | 0.225 | ||
| DGFRCNN-RS | ✓ | 0.111 | 0.229 | ||
| DGFRCNN-RS | ✓ | ✓ | 0.112 | 0.231 | |
| DGFRCNN-RS | ✓ | ✓ | ✓ | 0.115 | 0.234 |