Skip to content

Latest commit

 

History

History
180 lines (137 loc) · 11.6 KB

README_en.md

File metadata and controls

180 lines (137 loc) · 11.6 KB
Raw

English | 简体中文

PP-YOLOE-R

Content

Introduction

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.

Compared with PP-YOLOE, PP-YOLOE-R has made the following changes:

  • Rotated Task Alignment Learning
  • Decoupled Angle Prediction Head
  • Angle Prediction with DFL
  • Learnable Gating Unit for RepVGG
  • ProbIoU Loss

Model Zoo

Model Backbone mAP V100 TRT FP16 (FPS) RTX 2080 Ti TRT FP16 (FPS) Params (M) FLOPs (G) Lr Scheduler Angle Aug GPU Number images/GPU download config
PP-YOLOE-R-s CRN-s 73.82 114.5 69.8 8.09 43.46 3x oc RR 4 2 model config
PP-YOLOE-R-s CRN-s 79.42 114.5 69.8 8.09 43.46 3x oc MS+RR 4 2 model config
PP-YOLOE-R-m CRN-m 77.64 86.8 55.1 23.96 127.00 3x oc RR 4 2 model config
PP-YOLOE-R-m CRN-m 79.71 86.8 55.1 23.96 127.00 3x oc MS+RR 4 2 model config
PP-YOLOE-R-l CRN-l 78.14 69.7 48.3 53.29 281.65 3x oc RR 4 2 model config
PP-YOLOE-R-l CRN-l 80.02 69.7 48.3 53.29 281.65 3x oc MS+RR 4 2 model config
PP-YOLOE-R-x CRN-x 78.28 50.7 37.1 100.27 529.82 3x oc RR 4 2 model config
PP-YOLOE-R-x CRN-x 80.73 50.7 37.1 100.27 529.82 3x oc MS+RR 4 2 model config

Notes:

  • if GPU number or mini-batch size is changed, learning rate should be adjusted according to the formula lrnew = lrdefault * (batch_sizenew * GPU_numbernew) / (batch_sizedefault * GPU_numberdefault).
  • 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 to reproduce the results.

Getting Start

Refer to Data-Preparation to prepare data.

Training

Single GPU Training

CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml

Multiple GPUs Training

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml

Inference

Run the follow command to infer single image, the result of inference will be saved in output directory by default.

python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_img=demo/P0861__1.0__1154___824.png --draw_threshold=0.5

Evaluation on DOTA Dataset

Refering to DOTA Task, You need to submit a zip file containing results for all test images for evaluation. The detection results of each category are stored in a txt file, each line of which is in the following format image_id score x1 y1 x2 y2 x3 y3 x4 y4. To evaluate, you should submit the generated zip file to the Task1 of DOTA Evaluation. You can run the following command to get the inference results of test dataset:

python tools/infer.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams --infer_dir=/path/to/test/images --output_dir=output_ppyoloe_r --visualize=False --save_results=True

Process the prediction results into the format required for the official website evaluation:

python configs/rotate/tools/generate_result.py --pred_txt_dir=output_ppyoloe_r/ --output_dir=submit/ --data_type=dota10

zip -r submit.zip submit

Speed testing

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 the develop version. Using Paddle-TRT to test speed, run following command

# export inference model
python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True

# speed testing
CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode trt_fp16

Using Paddle to test speed, run following command

# export inference model
python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams

# speed testing
CUDA_VISIBLE_DEVICES=0 python configs/rotate/tools/inference_benchmark.py --model_dir output_inference/ppyoloe_r_crn_l_3x_dota/ --image_dir /path/to/dota/test/dir --run_mode paddle

Deployment

Using Paddle to for deployment, run following command

# export inference model
python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams

# inference single image
python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=paddle --device=gpu

Using Paddle-TRT for deployment, run following command

# export inference model
python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams trt=True

# inference single image
python deploy/python/infer.py --image_file demo/P0072__1.0__0___0.png --model_dir=output_inference/ppyoloe_r_crn_l_3x_dota --run_mode=trt_fp16 --device=gpu

Notes:

  • 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 the develop version

Using ONNX Runtime for deployment, run following command

# export inference model
python tools/export_model.py -c configs/rotate/ppyoloe_r/ppyoloe_r_crn_l_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_r_crn_l_3x_dota.pdparams export_onnx=True

# install paddle2onnx
pip install paddle2onnx

# convert to onnx model
paddle2onnx --model_dir output_inference/ppyoloe_r_crn_l_3x_dota --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 11 --save_file ppyoloe_r_crn_l_3x_dota.onnx

# inference single image
python configs/rotate/tools/onnx_infer.py --infer_cfg output_inference/ppyoloe_r_crn_l_3x_dota/infer_cfg.yml --onnx_file ppyoloe_r_crn_l_3x_dota.onnx --image_file demo/P0072__1.0__0___0.png

Appendix

Ablation experiments of PP-YOLOE-R

Model mAP Params(M) FLOPs(G)
Baseline 75.61 50.65 269.09
+Rotated Task Alignment Learning 77.24 50.65 269.09
+Decoupled Angle Prediction Head 77.78 52.20 272.72
+Angle Prediction with DFL 78.01 53.29 281.65
+Learnable Gating Unit for RepVGG 78.14 53.29 281.65

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},
  journal={arXiv preprint arXiv:2203.16250},
  year={2022}
}

@article{llerena2021gaussian,
  title={Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection},
  author={Llerena, Jeffri M and Zeni, Luis Felipe and Kristen, Lucas N and Jung, Claudio},
  journal={arXiv preprint arXiv:2106.06072},
  year={2021}
}