@@ -51,7 +51,7 @@ For labeled data, each image corresponds to a txt file with the same name, and e
x1 y1 x2 y2 x3 y3 x4 y4 class_name difficult
```
### Slicing data with single scale
#### Slicing data with single scale
The image resolution of DOTA dataset is relatively high, so we usually slice the images before training and testing. To slice the images with a single scale, you can use the command below
Rotated object detction uses the standard COCO data format, and you can convert your dataset to COCO format to train the model. The annotations of standard COCO format contains the following information
**It should be noted that `bbox` is the horizontal bouding box, and `segmentation` is four points of rotated bounding box (clockwise or counterclockwise). The `bbox` can be empty when training rotated object detector, and it is recommended to generate `bbox` according to `segmentation`**. In PaddleDetection 2.4 and earlier versions, `bbox` represents the rotated bounding box [x, y, w, h, angle] and `segmentation` is empty. **But this format is no longer supported after PaddleDetection 2.5, please download the latest dataset or convert to standard COCO format**.
## Installation
Models of rotated object detection depend on external operators for training, evaluation, etc. In Linux environment, you can execute the following command to compile and install.
title={PP-YOLOE-R: An Efficient Anchor-Free Rotated Object Detector},
author={Wang, Xinxin and Wang, Guanzhong and Dang, Qingqing and Liu, Yi and Hu, Xiaoguang and Yu, Dianhai},
journal={arXiv preprint arXiv:2211.02386},
year={2022}
}
@article{xu2022pp,
title={PP-YOLOE: An evolved version of YOLO},
author={Xu, Shangliang and Wang, Xinxin and Lv, Wenyu and Chang, Qinyao and Cui, Cheng and Deng, Kaipeng and Wang, Guanzhong and Dang, Qingqing and Wei, Shengyu and Du, Yuning and others},
PP-YOLOE-R is an efficient anchor-free rotated object detector. Based on PP-YOLOE, PP-YOLOE-R introduces a bag of useful tricks to improve detection precision at the expense of marginal parameters and computations.PP-YOLOE-R-l and PP-YOLOE-R-x achieve 78.14 and 78.27 mAP respectively on DOTA 1.0 dataset with single-scale training and testing, which outperform almost all other rotated object detectors. With multi-scale training and testing, the detection precision of PP-YOLOE-R-l and PP-YOLOE-R-x is further improved to 80.02 and 80.73 mAP. In this case, PP-YOLOE-R-x surpasses all anchor-free methods and demonstrates competitive performance to state-of-the-art anchor-based two-stage model. Moreover, PP-YOLOE-R-s and PP-YOLOE-R-m can achieve 79.42 and 79.71 mAP with multi-scale training and testing, which is an excellent result considering the parameters and GLOPS of these two models. While maintaining high precision, PP-YOLOE-R avoids using special operators, such as Deformable Convolution or Rotated RoI Align, to be deployed friendly on various hardware. At the input resolution of 1024$\times$1024, PP-YOLOE-R-s/m/l/x can reach 69.8/55.1/48.3/37.1 FPS on RTX 2080 Ti and 114.5/86.8/69.7/50.7 FPS on Tesla V100 GPU with TensorRT and FP16-precision. For more details, please refer to our technical report.
PP-YOLOE-R is an efficient anchor-free rotated object detector. Based on PP-YOLOE, PP-YOLOE-R introduces a bag of useful tricks to improve detection precision at the expense of marginal parameters and computations.PP-YOLOE-R-l and PP-YOLOE-R-x achieve 78.14 and 78.27 mAP respectively on DOTA 1.0 dataset with single-scale training and testing, which outperform almost all other rotated object detectors. With multi-scale training and testing, the detection precision of PP-YOLOE-R-l and PP-YOLOE-R-x is further improved to 80.02 and 80.73 mAP. In this case, PP-YOLOE-R-x surpasses all anchor-free methods and demonstrates competitive performance to state-of-the-art anchor-based two-stage model. Moreover, PP-YOLOE-R-s and PP-YOLOE-R-m can achieve 79.42 and 79.71 mAP with multi-scale training and testing, which is an excellent result considering the parameters and GLOPS of these two models. While maintaining high precision, PP-YOLOE-R avoids using special operators, such as Deformable Convolution or Rotated RoI Align, to be deployed friendly on various hardware. At the input resolution of 1024$\times$1024, PP-YOLOE-R-s/m/l/x can reach 69.8/55.1/48.3/37.1 FPS on RTX 2080 Ti and 114.5/86.8/69.7/50.7 FPS on Tesla V100 GPU with TensorRT and FP16-precision. For more details, please refer to our [**technical report**](https://arxiv.org/abs/2211.02386).
@@ -25,22 +25,23 @@ Compared with PP-YOLOE, PP-YOLOE-R has made the following changes:
-[ProbIoU Loss](https://arxiv.org/abs/2106.06072)
## Model Zoo
| Model | Backbone | mAP | V100 TRT FP16 (FPS) | RTX 2080 Ti TRT FP16 (FPS) | Lr Scheduler | Angle | Aug | GPU Number | images/GPU | download | config |
- if **GPU number** or **mini-batch size** is changed, **learning rate** should be adjusted according to the formula **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)**.
- Models in model zoo is trained and tested with single scale by default. If `MS` is indicated in the data augmentation column, it means that multi-scale training and multi-scale testing are used. If `RR` is indicated in the data augmentation column, it means that RandomRotate data augmentation is used for training.
- CRN denotes CSPRepResNet proposed in PP-YOLOE
- The parameters and GLOPs of PP-YOLOE-R are calculated after re-parameterization, and the resolution of the input image is 1024x1024
- Speed is calculated and averaged by testing 2000 images on the DOTA test dataset. Refer to [Speed testing](#Speed-testing) to reproduce the results.
## Getting Start
...
...
@@ -82,7 +83,7 @@ zip -r submit.zip submit
### Speed testing
To test speed, make sure that **the version of TensorRT is larger than 8.2 and the version of PaddlePaddle is larger than 2.4.0rc**. Using Paddle Inference with TensorRT to test speed, run following command
You can use Paddle mode or Paddle-TRT mode for speed testing. When using Paddle-TRT for speed testing, make sure that **the version of TensorRT is larger than 8.2 and the version of PaddlePaddle is larger than 2.4.0rc**. Using Paddle-TRT to test speed, run following command
- When using Paddle-TRT for speed testing, make sure that **the version of TensorRT is larger than 8.2 and the version of PaddlePaddle is larger than 2.4.0rc**
## Appendix
...
...
@@ -129,6 +141,13 @@ Ablation experiments of PP-YOLOE-R
## Citations
```
@article{wang2022pp,
title={PP-YOLOE-R: An Efficient Anchor-Free Rotated Object Detector},
author={Wang, Xinxin and Wang, Guanzhong and Dang, Qingqing and Liu, Yi and Hu, Xiaoguang and Yu, Dianhai},
journal={arXiv preprint arXiv:2211.02386},
year={2022}
}
@article{xu2022pp,
title={PP-YOLOE: An evolved version of YOLO},
author={Xu, Shangliang and Wang, Xinxin and Lv, Wenyu and Chang, Qinyao and Cui, Cheng and Deng, Kaipeng and Wang, Guanzhong and Dang, Qingqing and Wei, Shengyu and Du, Yuning and others},
