[WACV2021] Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors (arXiv)
Please cite the article in your publications if it helps your research:
@inproceedings{yi2021oriented,
title={Oriented object detection in aerial images with box boundary-aware vectors},
author={Yi, Jingru and Wu, Pengxiang and Liu, Bo and Huang, Qiaoying and Qu, Hui and Metaxas, Dimitris},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={2150--2159},
year={2021}
}
Oriented object detection in aerial images is a challenging task as the objects in aerial images are displayed in arbitrary directions and are usually densely packed. Current oriented object detection methods mainly rely on two-stage anchor-based detectors. However, the anchor-based detectors typically suffer from a severe imbalance issue between the positive and negative anchor boxes. To address this issue, in this work we extend the horizontal keypoint-based object detector to the oriented object detection task. In particular, we first detect the center keypoints of the objects, based on which we then regress the box boundary-aware vectors (BBAVectors) to capture the oriented bounding boxes. The box boundary-aware vectors are distributed in the four quadrants of a Cartesian coordinate system for all arbitrarily oriented objects. To relieve the difficulty of learning the vectors in the corner cases, we further classify the oriented bounding boxes into horizontal and rotational bounding boxes. In the experiment, we show that learning the box boundary-aware vectors is superior to directly predicting the width, height, and angle of an oriented bounding box, as adopted in the baseline method. Besides, the proposed method competes favorably with state-of-the-art methods.
Evaluation Results on DOTA-v1.0
When training the BBAVectors+rh on 4 RTX6000 GPUs with a larger batch size--batch_size 48, we get a higher mAP (75.36) than the reported mAP (72.32) in the paper. We add the result to our final version. We thank the public visitors for their effort. The model weights can be downloaded from the following links: GoogleDrive and Dropbox.
## model_50.pth
mAP: 0.7536283690546086
ap of each class: plane:0.8862514770737425, baseball-diamond:0.8406009896282075, bridge:0.521285610860641, ground-track-field:0.6955552280263699, small-vehicle:0.7825702607967113, large-vehicle:0.8040010247209182, ship:0.8805575982076236, tennis-court:0.9087489402165854, basketball-court:0.8722663525600673, storage-tank:0.8638699841268725, soccer-ball-field:0.5610545208583243, roundabout:0.6562139014619145, harbor:0.6709747110284013, swimming-pool:0.7208480121858474, helicopter:0.6396269240669054
## model_43.pth
mAP: 0.7492727335105831
ap of each class: plane:0.8859121197958046, baseball-diamond:0.8483251642688572, bridge:0.5214374843409882, ground-track-field:0.6560710395759289, small-vehicle:0.7773671634218439, large-vehicle:0.7427879633964128, ship:0.8804625721887132, tennis-court:0.908816372618596, basketball-court:0.862399364058993, storage-tank:0.8670730838290734, soccer-ball-field:0.5987801663737911, roundabout:0.6401450110418495, harbor:0.6698206063852568, swimming-pool:0.7071826121359568, helicopter:0.672510279226682Ubuntu 18.04, Python 3.6.10, PyTorch 1.6.0, OpenCV-Python 4.3.0.36
Download and install the DOTA development kit DOTA_devkit and put it under datasets folder.
Please uncomment the nn.BatchNorm2d(head_conv) in ctrbox_net.py to avoid NAN loss when training with a smaller batch size. Note that the current version of ctrbox_net.py matches the uploaded weights.
You mush split the DOTA images from DOTA_devkit before training, testing and evaluation.
The dota trainval and test datasets are cropped into 600×600 patches with a stride of 100 and two scales 0.5 and 1.
The trainval.txt and test.txt used in datasets/dataset_dota.py contain the list of image names without suffix, example:
P0000__0.5__0___0
P0000__0.5__0___1000
P0000__0.5__0___1500
P0000__0.5__0___2000
P0000__0.5__0___2151
P0000__0.5__0___500
P0000__0.5__1000___0
Some people would be interested in the format of the ground-truth, I provide some examples for DOTA dataset:
Format: x1, y1, x2, y2, x3, y3, x4, y4, category, difficulty
Examples:
275.0 463.0 411.0 587.0 312.0 600.0 222.0 532.0 tennis-court 0
341.0 376.0 487.0 487.0 434.0 556.0 287.0 444.0 tennis-court 0
428.0 6.0 519.0 66.0 492.0 108.0 405.0 50.0 bridge 0
python main.py --data_dir dataPath --epochs 80 --batch_size 16 --dataset dota --phase trainpython main.py --data_dir dataPath --batch_size 16 --dataset dota --phase testpython main.py --data_dir dataPath --conf_thresh 0.1 --batch_size 16 --dataset dota --phase evalYou may change conf_thresh to get a better mAP.
Please zip and upload the generated merge_dota for DOTA Task1 evaluation.