in modular design, and wealthy data augmentation methods, network components, loss functions, etc.
PaddleDetection supported practical projects such as industrial quality inspection, remote sensing
image object detection, and automatic inspection with its practical features such as model compression
and multi-platform deployment.
PaddleDetection is an end-to-end object detection development kit based on PaddlePaddle, which aims to help developers in the whole development of constructing, training, optimizing and deploying detection models in a faster and better way.
[PP-YOLO](https://arxiv.org/abs/2007.12099), which is faster and has higer performance than YOLOv4,
has been released, it reached mAP(0.5:0.95) as 45.2%(newest 45.9%) on COCO test2019 dataset and
72.9 FPS on single Test V100. Please refer to [PP-YOLO](configs/ppyolo/README.md) for details.
PaddleDetection implements varied mainstream object detection algorithms in modular design, and provides wealthy data augmentation methods, network components(such as backbones), loss functions, etc., and integrates abilities of model compression and cross-platform high-performance deployment.
**Now all models in PaddleDetection require PaddlePaddle version 1.8 or higher, or suitable develop version.**
After a long time of industry practice polishing, PaddleDetection has had smooth and excellent user experience, it has been widely used by developers in more than ten industries such as industrial quality inspection, remote sensing image object detection, automatic inspection, new retail, Internet, and scientific research.
<divalign="center">
<imgsrc="docs/images/000000570688.jpg"/>
<imgsrc="docs/images/football.gif"width='800'/>
</div>
## Introduction
Features:
- Rich models:
PaddleDetection provides rich of models, including 100+ pre-trained models
such as object detection, instance segmentation, face detection etc. It covers
the champion models, the practical detection models for cloud and edge device.
- Production Ready:
Key operations are implemented in C++ and CUDA, together with PaddlePaddle's
highly efficient inference engine, enables easy deployment in server environments.
- Highly Flexible:
Components are designed to be modular. Model architectures, as well as data
preprocess pipelines, can be easily customized with simple configuration
changes.
- Performance Optimized:
With the help of the underlying PaddlePaddle framework, faster training and
reduced GPU memory footprint is achieved. Notably, YOLOv3 training is
much faster compared to other frameworks. Another example is Mask-RCNN
(ResNet50), we managed to fit up to 4 images per GPU (Tesla V100 16GB) during
<aname="vd">[1]</a> [ResNet-vd](https://arxiv.org/pdf/1812.01187) models offer much improved accuracy with negligible performance cost.
**NOTE:** ✓ for config file and pretrain model provided in [Model Zoo](docs/MODEL_ZOO.md), ✗ for not provided but is supported generally.
More models:
- EfficientDet
- FCOS
- CornerNet-Squeeze
- YOLOv4
- PP-YOLO
More Backbones:
- DarkNet
- VGG
- GCNet
- CBNet
Advanced Features:
- [x] **Synchronized Batch Norm**
- [x] **Group Norm**
- [x] **Modulated Deformable Convolution**
- [x] **Deformable PSRoI Pooling**
- [x] **Non-local and GCNet**
**NOTE:** Synchronized batch normalization can only be used on multiple GPU devices, can not be used on CPU devices or single GPU device.
The following is the relationship between COCO mAP and FPS on Tesla V100 of representative models of each architectures and backbones.
### Product dynamic
- 2020.10.01: Added SOLOv2 as an instance segmentation model, which reached 38.6 FPS on a single Tesla V100, 38.8 mask AP on Coco-Val dataset, and inference speed increased by 24% and mAP by 2.4 percentage points.
- 2020.09.30: Released the mobile-side detection demo, and you can directly scan the code for installation experience.
- 2020.09.21-27: [Object detection 7 days of punching class] Hand in hand to teach you from the beginning to the advanced level, in-depth understanding of the object detection algorithm life. Join the course QQ group (1136406895) to study together :)
- 2020.07.24: [PP-YOLO](https://arxiv.org/abs/2007.12099), which is **the most practical** object detection model, was released, it deeply considers the double demands of industrial applications for accuracy and speed, and reached accuracy as 45.2% (the latest 45.9%) on COCO dataset, inference speed as 72.9 FPS on a single Test V100. Please refer to [PP-YOLO](https://arxiv.org/abs/2007.12099) for details.
- 2020.06.11: Publish 676 classes of large-scale server-side practical object detection models that are applicable to most application scenarios and can be used directly for prediction or for fine-tuning other tasks.
### Features
-**Rich Models**
PaddleDetection provides rich of models, including **100+ pre-trained models** such as **object detection**, **instance segmentation**, **face detection** etc. It covers a variety of **global competition champion** schemes.
-**Use Concisely**
Modular design, decouple each network component, developers easily build and try various detection models and optimization strategies, quickly get high-performance, customized algorithm.
-**Getting Through End to End**
From data augmentation, constructing models, training, compression, depolyment, get through end to end, and complete support for multi-architecture, multi-device deployment for **cloud and edge device**.
-**High Performance:**
Based on the high performance core of PaddlePaddle, advantages of training speed and memory occupation are obvious. Support FP16 training, support multi-machine training.
#### Overview of Kit Structures
<table>
<tbody>
<tralign="center"valign="bottom">
<td>
<b>Architectures</b>
</td>
<td>
<b>Backbones</b>
</td>
<td>
<b>Components</b>
</td>
<td>
<b>Data Augmentation</b>
</td>
</tr>
<trvalign="top">
<td>
<ul><li><b>Two-Stage Detection</b></li>
<ul>
<li>Faster RCNN</li>
<li>FPN</li>
<li>Cascade-RCNN</li>
<li>Libra RCNN</li>
<li>Hybrid Task RCNN</li>
<li>PSS-Det RCNN</li>
</ul>
</ul>
<ul><li><b>One-Stage Detection</b></li>
<ul>
<li>RetinaNet</li>
<li>YOLOv3</li>
<li>YOLOv4</li>
<li>PP-YOLO</li>
<li>SSD</li>
</ul>
</ul>
<ul><li><b>Anchor Free</b></li>
<ul>
<li>CornerNet-Squeeze</li>
<li>FCOS</li>
<li>TTFNet</li>
</ul>
</ul>
<ul>
<li><b>Instance Segmentation</b></li>
<ul>
<li>Mask RCNN</li>
<li>SOLOv2</li>
</ul>
</ul>
<ul>
<li><b>Face-Detction</b></li>
<ul>
<li>FaceBoxes</li>
<li>BlazeFace</li>
<li>BlazeFace-NAS</li>
</ul>
</ul>
</td>
<td>
<ul>
<li>ResNet(&vd)</li>
<li>ResNeXt(&vd)</li>
<li>SENet</li>
<li>Res2Net</li>
<li>HRNet</li>
<li>Hourglass</li>
<li>CBNet</li>
<li>GCNet</li>
<li>DarkNet</li>
<li>CSPDarkNet</li>
<li>VGG</li>
<li>MobileNetv1/v3</li>
<li>GhostNet</li>
<li>Efficientnet</li>
</ul>
</td>
<td>
<ul><li><b>Common</b></li>
<ul>
<li>Sync-BN</li>
<li>Group Norm</li>
<li>DCNv2</li>
<li>Non-local</li>
</ul>
</ul>
<ul><li><b>FPN</b></li>
<ul>
<li>BiFPN</li>
<li>BFP</li>
<li>HRFPN</li>
<li>ACFPN</li>
</ul>
</ul>
<ul><li><b>Loss</b></li>
<ul>
<li>Smooth-L1</li>
<li>GIoU/DIoU/CIoU</li>
<li>IoUAware</li>
</ul>
</ul>
<ul><li><b>Post-processing</b></li>
<ul>
<li>SoftNMS</li>
<li>MatrixNMS</li>
</ul>
</ul>
<ul><li><b>Speed</b></li>
<ul>
<li>FP16 training</li>
<li>Multi-machine training </li>
</ul>
</ul>
</td>
<td>
<ul>
<li>Resize</li>
<li>Flipping</li>
<li>Expand</li>
<li>Crop</li>
<li>Color Distort</li>
<li>Random Erasing</li>
<li>Mixup </li>
<li>Cutmix </li>
<li>Grid Mask</li>
<li>Auto Augment</li>
</ul>
</td>
</tr>
</td>
</tr>
</tbody>
</table>
#### Overview of Model Performance
The relationship between COCO mAP and FPS on Tesla V100 of representative models of each architectures and backbones.
<divalign="center">
<imgsrc="docs/images/map_fps.png"width=800/>
<imgsrc="docs/images/map_fps.png"/>
</div>
**NOTE:**
-`CBResNet` stands for `Cascade-Faster-RCNN-CBResNet200vd-FPN`, which has highest mAP on COCO as 53.3% in PaddleDetection models
-`Cascade-Faster-RCNN` stands for `Cascade-Faster-RCNN-ResNet50vd-DCN`, which has been optimized to 20 FPS inference speed when COCO mAP as 47.8%
- The enhanced `YOLOv3-ResNet50vd-DCN` is 10.6 absolute percentage points higher than paper on COCO mAP, and inference speed is nearly 70% faster than the darknet framework
- All these models can be get in [Model Zoo](#Model-Zoo)
The following is the relationship between COCO mAP and FPS on Tesla V100 of SOTA object detecters and PP-YOLO, which is faster and has better performance than YOLOv4, and reached mAP(0.5:0.95) as 45.2% on COCO test2019 dataset and 72.9 FPS on single Test V100. Please refer to [PP-YOLO](configs/ppyolo/README.md) for details.
-`CBResNet stands` for `Cascade-Faster-RCNN-CBResNet200vd-FPN`, which has highest mAP on COCO as 53.3%
-`Cascade-Faster-RCNN` stands for `Cascade-Faster-RCNN-ResNet50vd-DCN`, which has been optimized to 20 FPS inference speed when COCO mAP as 47.8% in PaddleDetection models
- The enhanced PaddleDetection model `YOLOv3-ResNet50vd-DCN` is 10.6 absolute percentage points higher than paper on COCO mAP, and inference speed is 61.3 fps, nearly 70% faster than the darknet framework.
All these models can be get in [Model Zoo](#ModelZoo)
## Tutorials
## Tutorials
### Get Started
-[Installation guide](docs/tutorials/INSTALL.md)
-[Quick start on small dataset](docs/tutorials/QUICK_STARTED.md)
-[Pretrained models for pedestrian detection](docs/featured_model/CONTRIB.md)
-[Pretrained models for vehicle detection](docs/featured_model/CONTRIB.md)
-[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
-[PP-YOLO](configs/ppyolo/README.md): PP-YOLO reeached mAP as 45.3% on COCO dataset，and 72.9 FPS on single Tesla V100
-[Best single model of Open Images 2019-Object Detction](docs/featured_model/champion_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%.
-[Large-scale practical object detection models](docs/featured_model/LARGE_SCALE_DET_MODEL_en.md): Large-scale practical server-side detection pretrained models with 676 categories are provided for most application scenarios, which can be used not only for direct inference but also finetuning on other datasets.
- Universal object detection
-[Model library and baselines](docs/MODEL_ZOO_cn.md)
-[Best single model of Open Images 2019-Object Detction](docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md)
## Updates
v0.4.0 was released at `07/2020`, add PP-YOLO, TTFNet, HTC, ACFPN, etc. And add BlaceFace face landmark detection model, add a series of optimized SSDLite models on mobile side, add data augmentations GridMask and RandomErasing, add Matrix NMS and EMA training, and improved ease of use, fix many known bugs, etc. Please refer to [change log](docs/CHANGELOG.md) for details.
## License
PaddleDetection is released under the [Apache 2.0 license](LICENSE).
## Updates
v0.4.0 was released at `05/2020`, add PP-YOLO, TTFNet, HTC, ACFPN, etc. And add BlaceFace face landmark detection model, add a series of optimized SSDLite models on mobile side, add data augmentations GridMask and RandomErasing, add Matrix NMS and EMA training, and improved ease of use, fix many known bugs, etc.
Please refer to [版本更新文档](docs/CHANGELOG.md) for details.
