Skip to content

Latest commit

 

History

History
518 lines (401 loc) · 41.2 KB

README_en.md

File metadata and controls

518 lines (401 loc) · 41.2 KB
Raw

简体中文 | English

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

Product Update

  • 🔥 2022.8.26：PaddleDetection releasesrelease/2.5 version

    • 🗳 Model features：

      • Release PP-YOLOE+: Increased accuracy by a maximum of 2.4% mAP to 54.9% mAP, 3.75 times faster model training convergence rate, and up to 2.3 times faster end-to-end inference speed; improved generalization for multiple downstream tasks
      • Release PicoDet-NPU model which supports full quantization deployment of models; add PicoDet layout analysis model
      • Release PP-TinyPose Plus. With 9.1% AP accuracy improvement in physical exercise, dance, and other scenarios, our PP-TinyPose Plus supports unconventional movements such as turning to one side, lying down, jumping, and high lifts

    • 🔮 Functions in different scenarios

      • Release the pedestrian analysis tool PP-Human v2. It introduces four new behavior recognition: fighting, telephoning, smoking, and trespassing. The underlying algorithm performance is optimized, covering three core algorithm capabilities: detection, tracking, and attributes of pedestrians. Our model provides end-to-end development and model optimization strategies for beginners and supports online video streaming input.
      • First release PP-Vehicle, which has four major functions: license plate recognition, vehicle attribute analysis (color, model), traffic flow statistics, and violation detection. It is compatible with input formats, including pictures, online video streaming, and video. And we also offer our users a comprehensive set of tutorials for customization.

    • 💡 Cutting-edge algorithms：

      • Covers YOLO family classic and latest models: YOLOv3, PP-YOLOE (a real-time high-precision object detection model developed by Baidu PaddlePaddle), and cutting-edge detection algorithms such as YOLOv4, YOLOv5, YOLOX, YOLOv6, and YOLOv7
      • Newly add high precision detection model based on ViT backbone network, with a 55.7% mAP accuracy on COCO dataset; newly add multi-object tracking model OC-SORT; newly add ConvNeXt backbone network.

    • 📋 Industrial applications: Newly add Smart Fitness, Fighting recognition, and Visitor Analysis.

  • 2022.3.24：PaddleDetection releasedrelease/2.4 version

    • Release high-performanace SOTA object detection model PP-YOLOE. It integrates cloud and edge devices and provides S/M/L/X versions. In particular, Verson L has the accuracy as 51.4% on COCO test 2017 dataset, inference speed as 78.1 FPS on a single Test V100. It supports mixed precision training, 33% faster than PP-YOLOv2. Its full range of multi-sized models can meet different hardware arithmetic requirements, and adaptable to server, edge-device GPU and other AI accelerator cards on servers.
    • Release ultra-lightweight SOTA object detection model PP-PicoDet Plus with 2% improvement in accuracy and 63% improvement in CPU inference speed. Add PicoDet-XS model with a 0.7M parameter, providing model sparsification and quantization functions for model acceleration. No specific post processing module is required for all the hardware, simplifying the deployment.
    • Release the real-time pedestrian analysis tool PP-Human. It has four major functions: pedestrian tracking, visitor flow statistics, human attribute recognition and falling detection. For falling detection, it is optimized based on real-life data with accurate recognition of various types of falling posture. It can adapt to different environmental background, light and camera angle.
    • Add YOLOX object detection model with nano/tiny/S/M/L/X. X version has the accuracy as 51.8% on COCO Val2017 dataset.

  • More releases

Brief Introduction

PaddleDetection is an end-to-end object detection development kit based on PaddlePaddle. Providing over 30 model algorithm and over 300 pre-trained models, it covers object detection, instance segmentation, keypoint detection, multi-object tracking. In particular, PaddleDetection offers high- performance & light-weight industrial SOTA models on servers and mobile devices, champion solution and cutting-edge algorithm. PaddleDetection provides various data augmentation methods, configurable network components, loss functions and other advanced optimization & deployment schemes. In addition to running through the whole process of data processing, model development, training, compression and deployment, PaddlePaddle also provides rich cases and tutorials to accelerate the industrial application of algorithm.

Features

  • Rich model library: PaddleDetection provides over 250 pre-trained models including object detection, instance segmentation, face recognition, multi-object tracking. It covers a variety of global competition champion schemes.
  • Simple to use: Modular design, decoupling each network component, easy for developers to build and try various detection models and optimization strategies, quick access to high-performance, customized algorithm.
  • Getting Through End to End: PaddlePaddle gets through end to end from data augmentation, constructing models, training, compression, depolyment. It also supports multi-architecture, multi-device deployment for cloud and edge device.
  • High Performance: Due to the high performance core, PaddlePaddle has clear advantages in training speed and memory occupation. It also supports FP16 training and multi-machine training.

Exchanges

  • If you have any question or suggestion, please give us your valuable input via GitHub Issues

    Welcome to join PaddleDetection user groups on WeChat (scan the QR code, add and reply "D" to the assistant)

Kit Structure

Architectures Backbones Components Data Augmentation
    Object Detection
    • Faster RCNN
    • FPN
    • Cascade-RCNN
    • PSS-Det
    • RetinaNet
    • YOLOv3
    • PP-YOLOv1/v2
    • PP-YOLO-Tiny
    • PP-YOLOE
    • PP-YOLOE+
    • YOLOX
    • SSD
    • CenterNet
    • FCOS
    • TTFNet
    • TOOD
    • GFL
    • PP-PicoDet
    • DETR
    • Deformable DETR
    • Swin Transformer
    • Sparse RCNN
    Instance Segmentation
    • Mask RCNN
    • Cascade Mask RCNN
    • SOLOv2
    Face Detection
    • BlazeFace
    Multi-Object-Tracking
    • JDE
    • FairMOT
    • DeepSORT
    • ByteTrack
    • OC-SORT
    KeyPoint-Detection
    • HRNet
    • HigherHRNet
    • Lite-HRNet
    • PP-TinyPose
Details
  • ResNet(&vd)
  • Res2Net(&vd)
  • CSPResNet
  • SENet
  • Res2Net
  • HRNet
  • Lite-HRNet
  • DarkNet
  • CSPDarkNet
  • MobileNetv1/v3
  • ShuffleNet
  • GhostNet
  • BlazeNet
  • DLA
  • HardNet
  • LCNet
  • ESNet
  • Swin-Transformer
  • ConvNeXt
  • Vision Transformer
Common
  • Sync-BN
  • Group Norm
  • DCNv2
  • EMA
KeyPoint
  • DarkPose
FPN
  • BiFPN
  • CSP-PAN
  • Custom-PAN
  • ES-PAN
  • HRFPN
Loss
  • Smooth-L1
  • GIoU/DIoU/CIoU
  • IoUAware
  • Focal Loss
  • CT Focal Loss
  • VariFocal Loss
Post-processing
  • SoftNMS
  • MatrixNMS
Speed
  • FP16 training
  • Multi-machine training
Details
  • Resize
  • Lighting
  • Flipping
  • Expand
  • Crop
  • Color Distort
  • Random Erasing
  • Mixup
  • AugmentHSV
  • Mosaic
  • Cutmix
  • Grid Mask
  • Auto Augment
  • Random Perspective

Model Performance

Performance comparison of Cloud models

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

Clarification：

  • ViT stands for ViT-Cascade-Faster-RCNN, which has highest mAP on COCO as 55.7%
  • Cascade-Faster-RCNNstands 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-YOLOE are optimized PP-YOLO v2. It reached accuracy as 51.4% on COCO dataset, inference speed as 78.1 FPS on Tesla V100
  • PP-YOLOE+ are optimized PP-YOLOE. It reached accuracy as 53.3% on COCO dataset, inference speed as 78.1 FPS on Tesla V100
  • The models in the figure are available in the model library
Performance omparison on mobiles

The comparison between COCO mAP and FPS on Qualcomm Snapdragon 865 processor of models on mobile devices.

Clarification：

  • Tests were conducted on Qualcomm Snapdragon 865 (4 *A77 + 4 *A55) batch_size=1, 4 thread, and NCNN inference library, test script see MobileDetBenchmark
  • PP-PicoDet and PP-YOLO-Tiny are self-developed models of PaddleDetection, and other models are not tested yet.

Model libraries

1. General detection

PP-YOLOE series Recommended scenarios: Cloud GPU such as Nvidia V100, T4 and edge devices such as Jetson series

Model COCO Accuracy（mAP） V100 TensorRT FP16 Speed(FPS) Configuration Download
PP-YOLOE+_s 43.9 333.3 link download
PP-YOLOE+_m 50.0 208.3 link download
PP-YOLOE+_l 53.3 149.2 link download
PP-YOLOE+_x 54.9 95.2 link download

PP-PicoDet series Recommended scenarios: Mobile chips and x86 CPU devices, such as ARM CPU(RK3399, Raspberry Pi) and NPU(BITMAIN)

Model COCO Accuracy（mAP） Snapdragon 865 four-thread speed (ms) Configuration Download
PicoDet-XS 23.5 7.81 Link Download
PicoDet-S 29.1 9.56 Link Download
PicoDet-M 34.4 17.68 Link Download
PicoDet-L 36.1 25.21 Link Download

Frontier detection algorithm

Model COCO Accuracy（mAP） V100 TensorRT FP16 speed(FPS) Configuration Download
YOLOX-l 50.1 107.5 Link Download
YOLOv5-l 48.6 136.0 Link Download

Other general purpose models doc

2. Instance segmentation
Model Introduction Recommended Scenarios COCO Accuracy(mAP) Configuration Download
Mask RCNN Two-stage instance segmentation algorithm
Edge-Cloud end
 box AP: 41.4
mask AP: 37.5		 Link Download
Cascade Mask RCNN Two-stage instance segmentation algorithm
Edge-Cloud end
 box AP: 45.7
mask AP: 39.7		 Link Download
SOLOv2 Lightweight single-stage instance segmentation algorithm
Edge-Cloud end
 mask AP: 38.0 Link Download
3. Keypoint detection
Model Introduction Recommended scenarios COCO Accuracy（AP） Speed Configuration Download
HRNet-w32 + DarkPose
Top-down Keypoint detection algorithm
Input size: 384x288
Edge-Cloud end
 78.3 T4 TensorRT FP16 2.96ms Link Download
HRNet-w32 + DarkPose Top-down Keypoint detection algorithm
Input size: 256x192		 Edge-Cloud end 78.0 T4 TensorRT FP16 1.75ms Link Download
PP-TinyPose Light-weight keypoint algorithm
Input size: 256x192		 Mobile 68.8 Snapdragon 865 four-thread 6.30ms Link Download
PP-TinyPose Light-weight keypoint algorithm
Input size: 128x96		 Mobile 58.1 Snapdragon 865 four-thread 2.37ms Link Download

Other keypoint detection models doc

4. Multi-object tracking PP-Tracking
Model Introduction Recommended scenarios Accuracy Configuration Download
ByteTrack SDE Multi-object tracking algorithm with detection model only Edge-Cloud end MOT-17 half val: 77.3 Link Download
FairMOT JDE multi-object tracking algorithm multi-task learning Edge-Cloud end MOT-16 test: 75.0 Link Download
OC-SORT SDE multi-object tracking algorithm with detection model only Edge-Cloud end MOT-16 half val: 75.5 Link -

Other multi-object tracking models docs

5. Industrial real-time pedestrain analysis tool-PP Human
Task End-to-End Speed（ms） Model Size
Pedestrian detection (high precision) 25.1ms Multi-object tracking 182M
Pedestrian detection (lightweight) 16.2ms Multi-object tracking 27M
Pedestrian tracking (high precision) 31.8ms Multi-object tracking 182M
Pedestrian tracking (lightweight) 21.0ms Multi-object tracking 27M
Attribute recognition (high precision) Single person8.5ms Object detection
Attribute recognition		 Object detection：182M
Attribute recognition：86M
Attribute recognition (lightweight) Single person 7.1ms Object detection
Attribute recognition		 Object detection：182M
Attribute recognition：86M
Falling detection Single person 10ms Multi-object tracking
Keypoint detection
Behavior detection based on key points		 Multi-object tracking：182M
Keypoint detection：101M
Behavior detection based on key points: 21.8M
Intrusion detection 31.8ms Multi-object tracking 182M
Fighting detection 19.7ms Video classification 90M
Smoking detection Single person 15.1ms Object detection
Object detection based on Human Id		 Object detection：182M
Object detection based on Human ID: 27M
Phoning detection Single person ms Object detection
Image classification based on Human ID		 Object detection：182M
Image classification based on Human ID：45M

Please refer to docs for details.

6. Industrial real-time vehicle analysis tool-PP Vehicle
Task End-to-End Speed（ms） Model Size
Vehicle detection (high precision) 25.7ms object detection 182M
Vehicle detection (lightweight) 13.2ms object detection 27M
Vehicle tracking (high precision) 40ms multi-object tracking 182M
Vehicle tracking (lightweight) 25ms multi-object tracking 27M
Plate Recognition 4.68ms plate detection
plate recognition		 Plate detection：3.9M
Plate recognition：12M
Vehicle attribute 7.31ms attribute recognition 7.2M

Please refer to docs for details.

Document tutorials

Introductory tutorials

Advanced tutorials

Courses

  • [Theoretical foundation] Object detection 7-day camp: Overview of object detection tasks, details of RCNN series object detection algorithm and YOLO series object detection algorithm, PP-YOLO optimization strategy and case sharing, introduction and practice of AnchorFree series algorithm

  • [Industrial application] AI Fast Track industrial object detection technology and application: Super object detection algorithms, real-time pedestrian analysis system PP-Human, breakdown and practice of object detection industrial application

  • [Industrial features] 2022.3.26 Smart City Industry Seven-Day Class : Urban planning, Urban governance, Smart governance service, Traffic management, community governance.

  • [Academic exchange] 2022.9.27 YOLO Vision Event: As the first YOLO-themed event, PaddleDetection was invited to communicate with the experts in the field of Computer Vision around the world.

Industrial tutorial examples

Applications

Recommended third-party tutorials

Version updates

Please refer to the Release note for more details about the updates

License

PaddlePaddle is provided under the Apache 2.0 license

Contribute your code

We appreciate your contributions and your feedback！

  • Thank Mandroide for code cleanup and
  • Thank FL77N for Sparse-RCNNmodel
  • Thank Chen-Song for Swin Faster-RCNNmodel
  • Thank yangyudong, hchhtc123 for developing PP-Tracking GUI interface
  • Thank Shigure19 for developing PP-TinyPose fitness APP
  • Thank manangoel99 for Wandb visualization methods

Quote

@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}
}