The oriented form of fcos.
@article{tian2019fcos,
title={FCOS: Fully Convolutional One-Stage Object Detection},
author={Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
journal={arXiv preprint arXiv:1904.01355},
year={2019}
}
In order to predict direction, I add an one channel convolution layer on the top of the regression features. I concatenate the 4-dim bbox predtions and 1-dim theta predictions as the final obb predictions. following original FCOS, I use the PolyIoULoss between the obb predictions and ground-truths as the loss of bboxes.
note: Cannot predict left, right, up, bottom targets with theta togather, because direction predictions cannot share same scale module with distance preditions.
Backbone | Data | Lr schd | centerness on reg | center sampling | norm on bbox | box AP |
---|---|---|---|---|---|---|
R50-FPN | DOTA1.0 | 1x | - | - | - | 72.1 |