未验证 提交 49ae43c5 编写于 作者: L littletomatodonkey 提交者: GitHub

Add detailed doc for PSS-DET. (#659)

* add detailed doc
* add zh_CN doc of pssdet
* fix link
上级 38d1517f
...@@ -123,6 +123,7 @@ ...@@ -123,6 +123,7 @@
- [车辆检测预训练模型](docs/featured_model/CONTRIB_cn.md) - [车辆检测预训练模型](docs/featured_model/CONTRIB_cn.md)
- [Objects365 2019 Challenge夺冠模型](docs/featured_model/CACascadeRCNN.md) - [Objects365 2019 Challenge夺冠模型](docs/featured_model/CACascadeRCNN.md)
- [Open Images 2019-Object Detction比赛最佳单模型](docs/featured_model/OIDV5_BASELINE_MODEL.md) - [Open Images 2019-Object Detction比赛最佳单模型](docs/featured_model/OIDV5_BASELINE_MODEL.md)
- [服务器端实用目标检测模型](configs/rcnn_enhance/README.md): V100上速度20FPS时,COCO mAP高达47.8%。
## 许可证书 ## 许可证书
......
...@@ -136,6 +136,7 @@ The following is the relationship between COCO mAP and FPS on Tesla V100 of repr ...@@ -136,6 +136,7 @@ The following is the relationship between COCO mAP and FPS on Tesla V100 of repr
- [YOLOv3 enhanced model](docs/featured_model/YOLOv3_ENHANCEMENT.md): Compared to MAP of 33.0% in paper, enhanced YOLOv3 reaches the MAP of 43.6%, and inference speed is improved as well - [YOLOv3 enhanced model](docs/featured_model/YOLOv3_ENHANCEMENT.md): Compared to MAP of 33.0% in paper, enhanced YOLOv3 reaches the MAP of 43.6%, and inference speed is improved as well
- [Objects365 2019 Challenge champion model](docs/featured_model/CACascadeRCNN.md) - [Objects365 2019 Challenge champion model](docs/featured_model/CACascadeRCNN.md)
- [Best single model of Open Images 2019-Object Detction](docs/featured_model/OIDV5_BASELINE_MODEL.md) - [Best single model of Open Images 2019-Object Detction](docs/featured_model/OIDV5_BASELINE_MODEL.md)
- [Practical Server-side detection method](configs/rcnn_enhance/README_en.md): Inference speed on single V100 GPU can reach 20FPS when COCO mAP is 47.8%.
## License ## License
......
## 服务器端实用目标检测方案(Practical Server-side detection, PSS-DET)
### 简介
* 近年来,学术界和工业界广泛关注图像中目标检测任务。基于[PaddleClas](https://github.com/PaddlePaddle/PaddleClas)中SSLD蒸馏方案训练得到的ResNet50_vd预训练模型(ImageNet1k验证集上Top1 Acc为82.39%),结合PaddleDetection中的丰富算子,飞桨提供了一种面向服务器端实用的目标检测方案PSS-DET(Practical Server Side Detection)。基于COCO2017目标检测数据集,V100单卡预测速度为为61FPS时,COCO mAP可达41.6%;预测速度为20FPS时,COCO mAP可达47.8%。
* 以标准的Faster RCNN ResNet50_vd FPN为例,下表给出了PSS-DET不同的模块的速度与精度收益。
| Trick | Train scale | Test scale | COCO mAP | Infer speed/FPS |
|- |:-: |:-: | :-: | :-: |
| `baseline` | 640x640 | 640x640 | 36.4% | 43.589 |
| +`test proposal=pre/post topk 500/300` | 640x640 | 640x640 | 36.2% | 52.512 |
| +`fpn channel=64` | 640x640 | 640x640 | 35.1% | 67.450 |
| +`ssld pretrain` | 640x640 | 640x640 | 36.3% | 67.450 |
| +`ciou loss` | 640x640 | 640x640 | 37.1% | 67.450 |
| +`DCNv2` | 640x640 | 640x640 | 39.4% | 60.345 |
| +`3x, multi-scale training` | 640x640 | 640x640 | 41.0% | 60.345 |
| +`auto augment` | 640x640 | 640x640 | 41.4% | 60.345 |
| +`libra sampling` | 640x640 | 640x640 | 41.6% | 60.345 |
基于该实验结论,PaddleDetection结合Cascade RCNN,使用更大的训练与评估尺度(1000x1500),最终在单卡V100上速度为20FPS,COCO mAP达47.8%。下图给出了目前类似速度的目标检测方法的速度与精度指标。
![pssdet](../../docs/images/pssdet.png)
**注意**
> 这里为了更方便地对比,统一将V100的预测耗时乘以1.2倍,近似转化为Titan V的预测耗时。
## 模型库
| 骨架网络 | 网络类型 | 每张GPU图片个数 | 学习率策略 |推理时间(fps) | Box AP | Mask AP | 下载 | 配置文件 |
| :---------------------- | :-------------: | :-------: | :-----: | :------------: | :----: | :-----: | :-------------: | :-----: |
| ResNet50-vd-FPN-Dcnv2 | Faster | 2 | 3x | 61.425 | 41.6 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_vd_fpn_3x_server_side.tar) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/master/configs/rcnn_server_side_det/faster_rcnn_dcn_r50_vd_fpn_3x_server_side.yml) |
| ResNet50-vd-FPN-Dcnv2 | Cascade Faster | 2 | 3x | 20.001 | 47.8 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r50_vd_fpn_3x_server_side.tar) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/master/configs/rcnn_server_side_det/cascade_rcnn_dcn_r50_vd_fpn_3x_server_side.yml) |
...@@ -2,7 +2,34 @@ ...@@ -2,7 +2,34 @@
## Introduction ## Introduction
* This is developed by PaddleDetection. Many useful tricks are utilized for the model training process. More details can be seen in the configuration file.
* In recent years, object detection tasks have attracted widespread attention. [PaddleClas](https://github.com/PaddlePaddle/PaddleClas) open-sourced the ResNet50_vd_SSLD pretrained model based on ImageNet(Top1 Acc 82.4%). And based on the pretrained model, PaddleDetection provided the PSS-DET (Practical Server-side detection) with the help of the rich operators in PaddleDetection. The inference speed can reach 61FPS on single V100 GPU when COCO mAP is 41.6%, and 20FPS when COCO mAP is 47.8%.
* We take the standard `Faster RCNN ResNet50_vd FPN` as an example. The following table shows ablation study of PSS-DET.
| Trick | Train scale | Test scale | COCO mAP | Infer speed/FPS |
|- |:-: |:-: | :-: | :-: |
| `baseline` | 640x640 | 640x640 | 36.4% | 43.589 |
| +`test proposal=pre/post topk 500/300` | 640x640 | 640x640 | 36.2% | 52.512 |
| +`fpn channel=64` | 640x640 | 640x640 | 35.1% | 67.450 |
| +`ssld pretrain` | 640x640 | 640x640 | 36.3% | 67.450 |
| +`ciou loss` | 640x640 | 640x640 | 37.1% | 67.450 |
| +`DCNv2` | 640x640 | 640x640 | 39.4% | 60.345 |
| +`3x, multi-scale training` | 640x640 | 640x640 | 41.0% | 60.345 |
| +`auto augment` | 640x640 | 640x640 | 41.4% | 60.345 |
| +`libra sampling` | 640x640 | 640x640 | 41.6% | 60.345 |
And the following figure shows `mAP-Speed` curves for some common detectors.
![pssdet](../../docs/images/pssdet.png)
**Note**
> For fair comparison, inference time for PSS-DET models on V100 GPU is transformed to Titan V GPU by multiplying by 1.2 times.
## Model Zoo ## Model Zoo
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册