Sedore is a object detection framework for rotated objects. This codebase is based on
Motivation of this repostiory is to provide a sustainable framework maintaining recent updates of mmdetection.
To adapt to object detection in aerial images, this repo has several unique and new features compared to the original mmdetection
-
Support Oriented Object Detection
In aerial images, objects are usually annotated by oriented bounding box (OBB). To support oriented object detection, we implement OBB Head (OBBRoIHead and OBBDenseHead). Also, we provide functions to transfer mask predictions to OBBs.
-
Cython Bbox Overlaps
Since one patch image with the size of 1024 × 1024 may contain over 1000 instances in DOTA, which make the bbox overlaps memroy consuming. To avoid out of GPU memory, we calculate the bbox overlaps in cython. The speed of cython version is close to the GPU version.
-
Rotation Augmentation
Since there are many orientation variations in aerial images, we implement the online rotation augmentation.
-
Rotated RoI Warping
Currently, we implement two types of rotated RoI Warping (Rotated RoI Align and Rotated Position Sensitive RoI Align).
This project is released under the Apache 2.0 license.
- Results are available in the Model zoo.
- You can find the detailed configs in configs/DOTA.
- The trained models are available at Google Drive or Baidu Drive.
Please refer to INSTALL.md for installation.
Please see GETTING_STARTED.md for the basic usage of mmdetection.
We appreciate all contributions to improve benchmarks for object detection in aerial images.