README_en.md

English | 简体中文

A High-Efficient Development Toolkit for Object Detection based on PaddlePaddle.

Latest News

• 🔥 2022.3.24：PaddleDetection release 2.4 version

  • Release GPU SOTA object detection series models (s/m/l/x) PP-YOLOE, achieving mAP as 51.4% on COCO test dataset and 78.1 FPS on Nvidia V100, supporting AMP training and its training speed is 33% faster than PP-YOLOv2.
  • Release enhanced models of PP-PicoDet, including PP-PicoDet-XS model with 0.7M parameters, its mAP promoted ~2% on COCO, inference speed accelerated 63% on CPU, and post-processing integrated into the network to optimize deployment pipeline.
  • Release real-time human analysis tool PP-Human, which is based on data from real-life situations, supporting pedestrian detection, attribute recognition, human tracking, multi-camera tracking, human statistics and action recognition.

• 2021.11.03: Release release/2.3 version. Release mobile object detection model ⚡PP-PicoDet, mobile keypoint detection model ⚡PP-TinyPose，Real-time tracking system PP-Tracking. Release object detection models, including Swin-Transformer, TOOD, GFL, release Sniper tiny object detection models and optimized PP-YOLO-EB model for EdgeBoard. Release mobile keypoint detection model Lite HRNet.

• 2021.08.10: Release release/2.2 version. Release Transformer object detection models, including DETR, Deformable DETR, Sparse RCNN. Release keypoint detection models, including DarkHRNet and model trained on MPII dataset. Release head-tracking and vehicle-tracking multi-object tracking models.

• 2021.05.20: Release release/2.1 version. Release Keypoint Detection, including HigherHRNet and HRNet, Multi-Object Tracking, including DeepSORT，JDE and FairMOT. Release model compression for PPYOLO series models.Update documents such as EXPORT ONNX MODEL.

Introduction

PaddleDetection is an end-to-end object detection development kit based on PaddlePaddle, which implements varied mainstream object detection, instance segmentation, tracking and keypoint detection algorithms in modular designwhich with configurable modules such as network components, data augmentations and losses, and release many kinds SOTA industry practice models, integrates abilities of model compression and cross-platform high-performance deployment, aims to help developers in the whole end-to-end development in a faster and better way.

PaddleDetection provides image processing capabilities such as object detection, instance segmentation, multi-object tracking, keypoint detection and etc.

PaddleDetection covers industrialization, smart city, security & protection, retail, medicare industry and etc.

Features

• Rich Models

  PaddleDetection provides rich of models, including 250+ pre-trained models such as object detection, instance segmentation, face detection, keypoint detection, multi-object tracking and etc, covering a variety of global competition champion schemes.

• Highly Flexible

  Components are designed to be modular. Model architectures, as well as data preprocess pipelines and optimization strategies, can be easily customized with simple configuration changes.

• Production Ready

  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. FP16 training and multi-machine training are supported as well.

Community

• If you have any problem or suggestion on PaddleDetection, please send us issues through GitHub Issues.

• Welcome to Join PaddleDetection QQ Group and Wechat Group (reply "Det").

  

Overview of Kit Structures

Architectures	Backbones	Components	Data Augmentation
Object Detection Faster RCNN FPN Cascade-RCNN Libra RCNN Hybrid Task RCNN PSS-Det RetinaNet YOLOv3 YOLOv4 PP-YOLOv1/v2 PP-YOLO-Tiny SSD CornerNet-Squeeze FCOS TTFNet PP-PicoDet DETR Deformable DETR Swin Transformer Sparse RCNN Instance Segmentation Mask RCNN SOLOv2 Face Detection FaceBoxes BlazeFace BlazeFace-NAS Multi-Object-Tracking JDE FairMOT DeepSort KeyPoint-Detection HRNet HigherHRNet	ResNet(&vd) ResNeXt(&vd) SENet Res2Net HRNet Hourglass CBNet GCNet DarkNet CSPDarkNet VGG MobileNetv1/v3 GhostNet Efficientnet BlazeNet	Common Sync-BN Group Norm DCNv2 Non-local KeyPoint DarkPose FPN BiFPN BFP HRFPN ACFPN Loss Smooth-L1 GIoU/DIoU/CIoU IoUAware Post-processing SoftNMS MatrixNMS Speed FP16 training Multi-machine training	Resize Lighting Flipping Expand Crop Color Distort Random Erasing Mixup Mosaic Cutmix Grid Mask Auto Augment Random Perspective

Overview of Model Performance

The relationship between COCO mAP and FPS on Tesla V100 of representative models of each server side architectures and backbones.

NOTE:

• 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

• PP-YOLO achieves mAP of 45.9% on COCO and 72.9FPS on Tesla V100. Both precision and speed surpass YOLOv4

• PP-YOLO v2 is optimized version of PP-YOLO which has mAP of 49.5% and 68.9FPS on Tesla V100

• All these models can be get in Model Zoo

The relationship between COCO mAP and FPS on Qualcomm Snapdragon 865 of representative mobile side models.

NOTE:

• All data tested on Qualcomm Snapdragon 865(4A77 + 4A55) processor with batch size of 1 and CPU threads of 4, and use NCNN library in testing, benchmark scripts is publiced at MobileDetBenchmark
• PP-PicoDet and PP-YOLO-Tiny are developed and released by PaddleDetection, other models are not provided in PaddleDetection.

Tutorials

Get Started

• Installation guide
• Prepare dataset
• Quick start on PaddleDetection

Advanced Tutorials

• Parameter configuration

  • Parameter configuration for RCNN model
  • Parameter configuration for PP-YOLO model

• Model Compression(Based on PaddleSlim)

  • Prune/Quant/Distill

• Inference and deployment

  • Export model for inference
  • Paddle Inference
    • Python inference
    • C++ inference
  • Paddle-Lite
  • Paddle Serving
  • Export ONNX model
  • Inference benchmark
  • Exporting to ONNX and using OpenVINO for inference

• Advanced development

  • New data augmentations
  • New detection algorithms

Model Zoo

• Universal object detection
  • Model library and baselines
  • PP-YOLOE
  • PP-YOLO
  • PP-PicoDet
  • Enhanced Anchor Free model--TTFNet
  • Mobile models
  • 676 classes of object detection
  • Two-stage practical PSS-Det
  • SSLD pretrained models
• Universal instance segmentation
  • SOLOv2
• Rotation object detection
  • S2ANet
• Keypoint detection
  • PP-TinyPose
  • HigherHRNet
  • HRNet
  • LiteHRNet
• Multi-Object Tracking
  • PP-Tracking
  • DeepSORT
  • JDE
  • FairMOT
• Vertical field
  • Face detection
  • Pedestrian detection
  • Vehicle detection
  • Real-Time Human Analysis Tool PP-Human
• Competition Plan
  • Objects365 2019 Challenge champion model
  • Best single model of Open Images 2019-Object Detection

Applications

• Christmas portrait automatic generation tool
• Android Fitness Demo

Updates

For the details of version update, please refer to Version Update Doc.

License

PaddleDetection is released under the Apache 2.0 license.

Contribution

Contributions are highly welcomed and we would really appreciate your feedback!!

• Thanks Mandroide for cleaning the code and unifying some function interface.
• Thanks FL77N for contributing the code of Sparse-RCNN model.
• Thanks Chen-Song for contributing the code of Swin Faster-RCNN model.
• Thanks yangyudong, hchhtc123 for contributing PP-Tracking GUI interface.
• Thanks Shigure19 for contributing PP-TinyPose fitness APP.

Citation

@misc{ppdet2019,
title={PaddleDetection, Object detection and instance segmentation toolkit based on PaddlePaddle.},
author={PaddlePaddle Authors},
howpublished = {\url{https://github.com/PaddlePaddle/PaddleDetection}},
year={2019}
}