PaddleDetection

Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.

12,914

2,904

12,914

1,252

View on GitHub

Top Related Projects

models

77,006

Models and examples built with TensorFlow

detectron2

30,784

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.

mmdetection

29,327

OpenMMLab Detection Toolbox and Benchmark

yolov5

51,450

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

Mask_RCNN

24,600

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow

darknet

21,700

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

Quick Overview

PaddleDetection is an end-to-end object detection toolkit based on PaddlePaddle, an open-source deep learning platform. It provides a rich set of object detection models, including YOLO, Faster R-CNN, and SSD, along with various data augmentation methods and training strategies. The project aims to make it easy for developers to train and deploy object detection models for various applications.

Pros

Comprehensive collection of state-of-the-art object detection models
Extensive data augmentation and training strategies for improved performance
Easy-to-use APIs and tools for model training, evaluation, and deployment
Good documentation and examples for various use cases

Cons

Primarily focused on PaddlePaddle ecosystem, which may limit integration with other deep learning frameworks
Steeper learning curve for users not familiar with PaddlePaddle
Some advanced features may require in-depth knowledge of object detection algorithms

Code Examples

Installing PaddleDetection:

pip install paddledetection

Training a YOLOv3 model:

from ppdet.engine import Trainer
from ppdet.core.workspace import load_config

cfg = load_config('configs/yolov3/yolov3_darknet53_270e_coco.yml')
trainer = Trainer(cfg)
trainer.train()

Performing inference with a trained model:

from ppdet.core.workspace import load_config
from ppdet.engine import Trainer

cfg = load_config('configs/yolov3/yolov3_darknet53_270e_coco.yml')
trainer = Trainer(cfg)
trainer.load_weights('output/yolov3_darknet53_270e_coco/model_final')
trainer.predict(['path/to/your/image.jpg'])

Getting Started

Install PaddlePaddle and PaddleDetection:

pip install paddlepaddle-gpu
pip install paddledetection

Clone the repository:

git clone https://github.com/PaddlePaddle/PaddleDetection.git
cd PaddleDetection

Train a model:

python tools/train.py -c configs/yolov3/yolov3_darknet53_270e_coco.yml

Evaluate the model:

python tools/eval.py -c configs/yolov3/yolov3_darknet53_270e_coco.yml -o weights=output/yolov3_darknet53_270e_coco/model_final

Perform inference:

python tools/infer.py -c configs/yolov3/yolov3_darknet53_270e_coco.yml -o weights=output/yolov3_darknet53_270e_coco/model_final --infer_img=demo/000000014439.jpg

Competitor Comparisons

models

77,006

Models and examples built with TensorFlow

Pros of TensorFlow Models

Broader scope, covering various ML tasks beyond object detection
Larger community and more extensive documentation
Better integration with TensorFlow ecosystem and tools

Cons of TensorFlow Models

Can be more complex to use due to its extensive feature set
May have slower development cycles for specific tasks like object detection
Potentially higher resource requirements for some models

Code Comparison

PaddleDetection:

from ppdet.core.workspace import create
from ppdet.engine import Trainer

model = create('YOLOv3')
trainer = Trainer(model=model, use_gpu=True)
trainer.train()

TensorFlow Models:

import tensorflow as tf
from object_detection import model_lib_v2

model_fn = model_lib_v2.get_model_fn(
    num_classes=90, pipeline_config_path='path/to/config')
estimator = tf.estimator.Estimator(model_fn=model_fn)
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)

Both repositories offer robust object detection capabilities, but PaddleDetection focuses specifically on this task, while TensorFlow Models covers a broader range of machine learning applications. PaddleDetection may be easier to use for beginners in object detection, while TensorFlow Models provides more flexibility for advanced users and diverse ML projects.

detectron2

30,784

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.

Pros of Detectron2

More extensive documentation and tutorials
Larger community and ecosystem of extensions
Better integration with PyTorch ecosystem

Cons of Detectron2

Steeper learning curve for beginners
Less focus on mobile and edge deployment
Fewer pre-trained models for specialized tasks

Code Comparison

PaddleDetection:

from ppdet.core.workspace import create
from ppdet.engine import Trainer

model = create('YOLOv3')
trainer = Trainer(model=model, use_gpu=True)
trainer.train()

Detectron2:

from detectron2.config import get_cfg
from detectron2.engine import DefaultTrainer

cfg = get_cfg()
cfg.merge_from_file("config.yaml")
trainer = DefaultTrainer(cfg)
trainer.resume_or_load(resume=False)
trainer.train()

Both frameworks offer similar high-level APIs for model creation and training. PaddleDetection uses a custom workspace system, while Detectron2 relies on a configuration-based approach. Detectron2's code is more tightly integrated with PyTorch conventions, while PaddleDetection uses PaddlePaddle-specific constructs.

mmdetection

29,327

OpenMMLab Detection Toolbox and Benchmark

Pros of mmdetection

More extensive model zoo with a wider variety of pre-trained models
Better documentation and community support
More flexible and modular architecture for easier customization

Cons of mmdetection

Steeper learning curve for beginners
Slightly more complex configuration system
May have higher computational requirements for some models

Code Comparison

mmdetection:

from mmdet.apis import init_detector, inference_detector

config_file = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
checkpoint_file = 'checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'
model = init_detector(config_file, checkpoint_file, device='cuda:0')
result = inference_detector(model, 'test.jpg')

PaddleDetection:

from ppdet.engine import Detector
from ppdet.utils.visualizer import visualize_results

model = Detector('faster_rcnn_r50_fpn_1x_coco')
result = model.predict('test.jpg')
visualize_results(result, 'test.jpg', output_dir='output')

Both repositories offer powerful object detection frameworks, but mmdetection provides more flexibility and a larger model zoo, while PaddleDetection offers a simpler API for quick deployment.

yolov5

51,450

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

Pros of YOLOv5

Simpler architecture and easier to understand for beginners
Faster training and inference times
More extensive documentation and community support

Cons of YOLOv5

Limited flexibility for customization compared to PaddleDetection
Fewer pre-trained models and datasets available
Less support for advanced features like multi-object tracking

Code Comparison

YOLOv5:

from yolov5 import YOLOv5

model = YOLOv5('yolov5s.pt')
results = model('image.jpg')
results.show()

PaddleDetection:

from ppdet.engine import Trainer
from ppdet.core.workspace import load_config

cfg = load_config('configs/yolov3/yolov3_darknet53_270e_coco.yml')
trainer = Trainer(cfg)
trainer.train()

YOLOv5 offers a more straightforward API for quick implementation, while PaddleDetection provides a more comprehensive framework with greater customization options. YOLOv5's code is more concise and easier to use out-of-the-box, whereas PaddleDetection requires more setup but offers more flexibility for advanced users.

Mask_RCNN

24,600

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow

Pros of Mask_RCNN

Simpler implementation, easier to understand and modify
Well-documented with detailed explanations and examples
Specifically designed for instance segmentation tasks

Cons of Mask_RCNN

Less frequently updated compared to PaddleDetection
Limited to Mask R-CNN architecture, while PaddleDetection offers multiple models
Smaller community and fewer contributions

Code Comparison

Mask_RCNN:

import mrcnn.model as modellib

model = modellib.MaskRCNN(mode="inference", config=config, model_dir=MODEL_DIR)
model.load_weights(COCO_MODEL_PATH, by_name=True)
results = model.detect([image], verbose=1)

PaddleDetection:

from ppdet.engine import Trainer
from ppdet.core.workspace import load_config, merge_config

cfg = load_config("configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml")
trainer = Trainer(cfg, mode='eval')
trainer.load_weights(cfg.weights)
trainer.predict([image])

Both repositories provide implementations for object detection and instance segmentation. Mask_RCNN focuses specifically on the Mask R-CNN architecture, while PaddleDetection offers a wider range of models and features. PaddleDetection is more actively maintained and has a larger community, but Mask_RCNN may be easier to understand and modify for those specifically interested in Mask R-CNN implementation.

darknet

21,700

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

Pros of darknet

Lightweight and fast, optimized for real-time object detection
Supports both CPU and GPU computation
Extensive documentation and community support

Cons of darknet

Limited to YOLO-based architectures
Less flexibility in model customization
Steeper learning curve for beginners

Code Comparison

darknet:

layer make_convolutional_layer(int batch, int h, int w, int c, int n, int groups, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor, int adam)
{
    layer l = {0};
    l.type = CONVOLUTIONAL;
    // ... (additional initialization)
}

PaddleDetection:

class ConvBNLayer(nn.Layer):
    def __init__(self,
                 ch_in,
                 ch_out,
                 filter_size=3,
                 stride=1,
                 groups=1,
                 padding=0,
                 act=None):
        super(ConvBNLayer, self).__init__()
        # ... (additional initialization)

PaddleDetection offers a more Pythonic and object-oriented approach, while darknet uses a C-style structure initialization. PaddleDetection provides greater flexibility in model architecture design, whereas darknet focuses on efficiency and simplicity for YOLO-based models.

Convert designs to code with AI

Introducing Visual Copilot: A new AI model to turn Figma designs to high quality code using your components.

Try Visual Copilot

README

ç®ä½ä¸æ | English

ðç®å½

ðç®å½
ðç®ä»
ð£ææ°è¿å±
â¡ï¸å¿«éå¼å§
ð¥ä½ä»£ç å¨æµç¨å¼å
ð«å¼æºç¤¾åº
â¨ä¸»è¦ç¹æ§
ð±å®è£
ð¥æç¨
ðFAQ
ð§©æ¨¡åç»ä»¶
ð±æ¨¡ååº
âï¸æ¨¡åæ§è½å¯¹æ¯
- ð¥ï¸æå¡å¨ç«¯æ¨¡åæ§è½å¯¹æ¯
- âï¸ç§»å¨ç«¯æ¨¡åæ§è½å¯¹æ¯
ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·
ð¡äº§ä¸å®è·µèä¾
ðä¼ä¸åºç¨æ¡ä¾
ðè®¸å¯è¯ä¹¦
ðå¼ç¨

ðç®ä»

ä¸»è¦æ¨¡åææç¤ºä¾å¦ä¸ï¼ç¹å»æ é¢å¯å¿«éè·³è½¬ï¼ï¼

éç¨ç®æ æ£æµ	å°ç®æ æ£æµ	æè½¬æ¡æ£æµ	3Dç®æ ç©æ£æµ

äººè¸æ£æµ	2Då³é®ç¹æ£æµ	å¤ç®æ è¿½è¸ª	å®ä¾åå²

è½¦è¾åæââè½¦çè¯å«	è½¦è¾åæââè½¦æµç»è®¡	è½¦è¾åæââè¿ç« æ£æµ	è½¦è¾åæââå±æ§åæ

è¡äººåæââé¯å¥åæ	è¡äººåæââè¡ä¸ºåæ	è¡äººåæââå±æ§åæ	è¡äººåæââäººæµç»è®¡

ð£ææ°è¿å±

ð¥2024.10.1 æ·»å ç®æ æ£æµãå®ä¾åå²é¢åä¸ç«å¼å¨æµç¨å¼åè½å:
- é£æ¡¨ä½ä»£ç å¼åå·¥å·PaddleXï¼ä¾æäºPaddleDetectionçåè¿ææ¯ï¼æ¯æäºç®æ æ£æµé¢åçä¸ç«å¼å¨æµç¨å¼åè½åï¼
  - ð¨ æ¨¡åä¸°å¯ä¸é®è°ç¨ï¼å°éç¨ç®æ æ£æµãå°ç®æ æ£æµåå®ä¾åå²æ¶åç55ä¸ªæ¨¡åæ´åä¸º3æ¡æ¨¡åäº§çº¿ï¼éè¿æç®ç**Python APIä¸é®è°ç¨ï¼å¿«éä½éªæ¨¡åææãæ¤å¤ï¼åä¸å¥APIï¼ä¹æ¯æå¾ååç±»ãå¾ååå²ãææ¬å¾åæºè½åæãéç¨OCRãæ¶åºé¢æµçå±è®¡200+æ¨¡åï¼å½¢æ20+ååè½æ¨¡åï¼æ¹ä¾¿å¼åèè¿è¡æ¨¡åç»åä½¿ç¨**ã
  - ð æé«æçéä½é¨æ§ï¼æä¾åºäºç»ä¸å½ä»¤åå¾å½¢çé¢ä¸¤ç§æ¹å¼ï¼å®ç°æ¨¡åç®æ´é«æçä½¿ç¨ãç»åä¸å®å¶ãæ¯æé«æ§è½é¨ç½²ãæå¡åé¨ç½²åç«¯ä¾§é¨ç½²çå¤ç§é¨ç½²æ¹å¼ãæ¤å¤ï¼å¯¹äºåç§ä¸»æµç¡¬ä»¶å¦è±ä¼è¾¾GPUãæä»è¯ãæè¾ãå¯æ¦çºªåæµ·åçï¼è¿è¡æ¨¡åå¼åæ¶ï¼é½å¯ä»¥**æ ç¼åæ¢**ã
- æ·»å å®ä¾åå²SOTAæ¨¡åMask-RT-DETR

RT-DETRè§£è¯»æç« ä¼ éé¨ï¼
- ãè¶è¶YOLOv8ï¼é£æ¡¨æ¨åºç²¾åº¦æé«çå®æ¶æ£æµå¨RT-DETRï¼ã
ä»£ç ä¼ éé¨ï¼RT-DETR

â¡ï¸å¿«éå¼å§

ð¥ä½ä»£ç å¨æµç¨å¼å

ð«å¼æºç¤¾åº

ðï¸ç¤¾åºè´¡ç®ï¼ PaddleDetectionéå¸¸æ¬¢è¿ä½ å å¥å°é£æ¡¨ç¤¾åºçå¼æºå»ºè®¾ä¸ï¼åä¸è´¡ç®æ¹å¼å¯ä»¥åèå¼æºé¡¹ç®å¼åæåã
ðç¤¾åºè¿ææ´»å¨

â¨ä¸»è¦ç¹æ§

ð§©æ¨¡ååè®¾è®¡

ð±ä¸°å¯çæ¨¡ååº

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·

PaddleDetectionæé äº§ä¸çº§ç¹è²æ¨¡åä»¥ååæå·¥å·ï¼PP-YOLOE+ãPP-PicoDetãPP-TinyPoseãPP-HumanV2ãPP-Vehicleçï¼éå¯¹éç¨ãé«é¢åç±»åºç¨åºæ¯æä¾æ·±åº¦ä¼åè§£å³æ¹æ¡ä»¥åé«åº¦éæçåæå·¥å·ï¼éä½å¼åèçè¯éãéæ©ææ¬ï¼éå¯¹ä¸å¡åºæ¯å¿«éåºç¨è½å°ãä¼ éé¨ï¼ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·ã

ð¡ðäº§ä¸çº§é¨ç½²å®è·µ

ð±å®è£

åèå®è£è¯´æè¿è¡å®è£ã

ð¥æç¨

æ·±åº¦å¦ä¹ å¥é¨æç¨

å¿«éå¼å§

æ°æ®åå¤

éç½®æä»¶è¯´æ

æ¨¡åå¼å

æ°å¢æ£æµæ¨¡å
äºæ¬¡å¼å

é¨ç½²æ¨ç

ðFAQ

FAQ/å¸¸è§é®é¢æ±æ»

ð§©æ¨¡åç»ä»¶

Backbones

Necks

Loss

Common

Data Augmentation

Post-processing

Training

Common

ð±æ¨¡ååº

2D Detection

Multi Object Tracking

KeyPoint Detection

Others

Instance Segmentation

Face Detection

BlazeFace

Semi-Supervised Detection

DenseTeacher

3D Detection

Vehicle Analysis Toolbox

PP-Vehicle

Human Analysis Toolbox

Sport Analysis Toolbox

PP-Sports

âï¸æ¨¡åæ§è½å¯¹æ¯

ð¥ï¸æå¡å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

æµè¯è¯´æ(ç¹å»å±å¼)

ViTä¸ºViT-Cascade-Faster-RCNNæ¨¡åï¼COCOæ°æ®émAPé«è¾¾55.7%
Cascade-Faster-RCNNä¸ºCascade-Faster-RCNN-ResNet50vd-DCNï¼PaddleDetectionå°å¶ä¼åå°COCOæ°æ®mAPä¸º47.8%æ¶æ¨çéåº¦ä¸º20FPS
PP-YOLOEæ¯å¯¹PP-YOLO v2æ¨¡åçè¿ä¸æ¥ä¼åï¼Lçæ¬å¨COCOæ°æ®émAPä¸º51.6%ï¼Tesla V100é¢æµéåº¦78.1FPS
PP-YOLOE+æ¯å¯¹PPOLOEæ¨¡åçè¿ä¸æ¥ä¼åï¼Lçæ¬å¨COCOæ°æ®émAPä¸º53.3%ï¼Tesla V100é¢æµéåº¦78.1FPS
YOLOXåYOLOv5åä¸ºåºäºPaddleDetectionå¤ç°ç®æ³ï¼YOLOv5ä»£ç å¨PaddleYOLOä¸ï¼åç§PaddleYOLO_MODEL
å¾ä¸æ¨¡ååå¯å¨ð±æ¨¡ååºä¸è·å

âï¸ç§»å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

æµè¯è¯´æ(ç¹å»å±å¼)

æµè¯æ°æ®åä½¿ç¨é«ééªé¾865(4xA77+4xA55)å¤çå¨ï¼batch sizeä¸º1, å¼å¯4çº¿ç¨æµè¯ï¼æµè¯ä½¿ç¨NCNNé¢æµåºï¼æµè¯èæ¬è§MobileDetBenchmark
PP-PicoDetåPP-YOLO-Tinyä¸ºPaddleDetectionèªç æ¨¡åï¼å¯å¨ð±æ¨¡ååºä¸è·åï¼å¶ä½æ¨¡åPaddleDetectionææªæä¾

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·

ðPP-YOLOE é«ç²¾åº¦ç®æ æ£æµæ¨¡å

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-YOLOEè¯´æã

ä¼ éé¨ï¼arXivè®ºæã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

æ¨¡ååç§°	COCOç²¾åº¦ï¼mAPï¼	V100 TensorRT FP16éåº¦(FPS)	æ¨èé¨ç½²ç¡¬ä»¶	éç½®æä»¶	æ¨¡åä¸è½½
PP-YOLOE+_l	53.3	149.2	æå¡å¨	é¾æ¥	ä¸è½½å°å

ä¼ éé¨ï¼å¨é¨é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
åä¸	åä½ç©æ£æµ	ç¨äºè¡èæ ½å¹ä¸åºäºå¾åççæµåç°åºæºå¨äººææ¯ï¼æä¾äºæ¥èª5ç§ä¸åè¡èåç§çå®å°å®ä¾	PP-YOLOE+ ä¸æ¸¸ä»»å¡	ä¸è½½é¾æ¥
éç¨	ä½ååºæ¯æ£æµ	ä½åæ°æ®éä½¿ç¨ExDarkï¼åæ¬ä»æä½åç¯å¢å°æ®åç¯å¢ç10ç§ä¸ååç§æ¡ä»¶ä¸çå¾çã	PP-YOLOE+ ä¸æ¸¸ä»»å¡	ä¸è½½é¾æ¥
å·¥ä¸	PCBçµè·¯æ¿ççµæ£æµ	å·¥ä¸æ°æ®éä½¿ç¨PKU-Market-PCBï¼è¯¥æ°æ®éç¨äºå°å·çµè·¯æ¿ï¼PCBï¼çççµæ£æµï¼æä¾äº6ç§å¸¸è§çPCBç¼ºé·	PP-YOLOE+ ä¸æ¸¸ä»»å¡	ä¸è½½é¾æ¥

ðPP-YOLOE-R é«æ§è½æè½¬æ¡æ£æµæ¨¡å

ç®ä»(ç¹å»å±å¼)

PP-YOLOE-Ræ¯ä¸ä¸ªé«æçåé¶æ®µAnchor-freeæè½¬æ¡æ£æµæ¨¡åï¼åºäºPP-YOLOE+å¼å¥äºä¸ç³»åæ¹è¿çç¥æ¥æåæ£æµç²¾åº¦ãæ ¹æ®ä¸åçç¡¬ä»¶å¯¹ç²¾åº¦åéåº¦çè¦æ±ï¼PP-YOLOE-Råå«s/m/l/xåä¸ªå°ºå¯¸çæ¨¡åãå¨DOTA 1.0æ°æ®éä¸ï¼PP-YOLOE-R-låPP-YOLOE-R-xå¨åå°ºåº¦è®ç»åæµè¯çæåµä¸åå«è¾¾å°äº78.14mAPå78.28 mAPï¼è¿å¨åå°ºåº¦è¯ä¼°ä¸è¶è¶äºå ä¹ææçæè½¬æ¡æ£æµæ¨¡åãéè¿å¤å°ºåº¦è®ç»åæµè¯ï¼PP-YOLOE-R-låPP-YOLOE-R-xçæ£æµç²¾åº¦è¿ä¸æ¥æåè³80.02mAPå80.73 mAPï¼è¶è¶äºææçAnchor-freeæ¹æ³å¹¶ä¸åæåè¿çAnchor-basedçä¸¤é¶æ®µæ¨¡åç²¾åº¦å ä¹ç¸å½ãå¨ä¿æé«ç²¾åº¦çåæ¶ï¼PP-YOLOE-Ré¿åä½¿ç¨ç¹æ®çç®åï¼ä¾å¦Deformable ConvolutionæRotated RoI Alignï¼ä½¿å¶è½è½»æ¾å°é¨ç½²å¨å¤ç§å¤æ ·çç¡¬ä»¶ä¸ã

ä¼ éé¨ï¼PP-YOLOE-Rè¯´æã

ä¼ éé¨ï¼arXivè®ºæã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

æ¨¡å	Backbone	mAP	V100 TRT FP16 (FPS)	RTX 2080 Ti TRT FP16 (FPS)	Params (M)	FLOPs (G)	å¦ä¹ ççç¥	è§åº¦è¡¨ç¤º	æ°æ®å¢å¹¿	GPUæ°ç®	æ¯GPUå¾çæ°ç®	æ¨¡åä¸è½½	éç½®æä»¶
PP-YOLOE-R-l	CRN-l	80.02	69.7	48.3	53.29	281.65	3x	oc	MS+RR	4	2	model	config

ä¼ éé¨ï¼å¨é¨é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
éç¨	æè½¬æ¡æ£æµ	ææææä½ ä¸æPP-YOLOE-Ræè½¬æ¡æ£æµï¼10åéå°èæ±æ°æ®éç²¾åº¦è®ç»è³95mAP	åºäºPP-YOLOE-Rçæè½¬æ¡æ£æµ	ä¸è½½é¾æ¥

ðPP-YOLOE-SOD é«ç²¾åº¦å°ç®æ æ£æµæ¨¡å

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-YOLOE-SOD å°ç®æ æ£æµæ¨¡åã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

- VisDroneæ°æ®éé¢è®ç»æ¨¡å

æ¨¡å	COCOAPI mAP^val 0.5:0.95	COCOAPI mAP^val 0.5	COCOAPI mAP^{test_dev 0.5:0.95}	COCOAPI mAP^test_dev 0.5	MatlabAPI mAP^{test_dev 0.5:0.95}	MatlabAPI mAP^test_dev 0.5	ä¸è½½	éç½®æä»¶
PP-YOLOE+_SOD-l	31.9	52.1	25.6	43.5	30.25	51.18	ä¸è½½é¾æ¥	éç½®æä»¶

ä¼ éé¨ï¼å¨é¨é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
éç¨	å°ç®æ æ£æµ	åºäºPP-YOLOE-SODçæ äººæºèªæå¾åæ£æµæ¡ä¾å¨æµç¨å®æã	åºäºPP-YOLOE-SODçæ äººæºèªæå¾åæ£æµ	ä¸è½½é¾æ¥

ð«PP-PicoDet è¶è½»éå®æ¶ç®æ æ£æµæ¨¡å

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-PicoDetè¯´æã

ä¼ éé¨ï¼arXivè®ºæã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

æ¨¡ååç§°	COCOç²¾åº¦ï¼mAPï¼	éªé¾865 åçº¿ç¨éåº¦(FPS)	æ¨èé¨ç½²ç¡¬ä»¶	éç½®æä»¶	æ¨¡åä¸è½½
PicoDet-L	36.1	39.7	ç§»å¨ç«¯ãåµå¥å¼	é¾æ¥	ä¸è½½å°å

ä¼ éé¨ï¼å¨é¨é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
æºæ§åå¸	éè·¯åå¾æ£æµ	éè¿å¨å¸æ¿ç¯å«è½¦è¾ä¸å®è£æåå¤´å¯¹è·¯é¢åå¾æ£æµå¹¶åæï¼å®ç°å¯¹è·¯é¢éæçåå¾è¿è¡çæ§ï¼è®°å½å¹¶éç¥ç¯å«äººåæ¸çï¼å¤§å¤§æåäºç¯å«äººæã	åºäºPP-PicoDetçè·¯é¢åå¾æ£æµ	ä¸è½½é¾æ¥

ð¡PP-Tracking å®æ¶å¤ç®æ è·è¸ªç³»ç»

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-Trackingè¯´æã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

æ¨¡ååç§°	æ¨¡åç®ä»	ç²¾åº¦	éåº¦(FPS)	æ¨èé¨ç½²ç¡¬ä»¶	éç½®æä»¶	æ¨¡åä¸è½½
ByteTrack	SDEå¤ç®æ è·è¸ªç®æ³ ä»åå«æ£æµæ¨¡å	MOT-17 test: 78.4	-	æå¡å¨ãç§»å¨ç«¯ãåµå¥å¼	é¾æ¥	ä¸è½½å°å
FairMOT	JDEå¤ç®æ è·è¸ªç®æ³ å¤ä»»å¡èåå¦ä¹ æ¹æ³	MOT-16 test: 75.0	-	æå¡å¨ãç§»å¨ç«¯ãåµå¥å¼	é¾æ¥	ä¸è½½å°å
OC-SORT	SDEå¤ç®æ è·è¸ªç®æ³ ä»åå«æ£æµæ¨¡å	MOT-17 half val: 75.5	-	æå¡å¨ãç§»å¨ç«¯ãåµå¥å¼	é¾æ¥	ä¸è½½å°å

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
éç¨	å¤ç®æ è·è¸ª	å¿«éä¸æåéå¤´ãå¤éå¤´è·è¸ª	PP-Trackingä¹æææç©è½¬å¤ç®æ è·è¸ª	ä¸è½½é¾æ¥

â·ï¸PP-TinyPose äººä½éª¨éª¼å³é®ç¹è¯å«

ç®ä»(ç¹å»å±å¼)

PaddleDetection ä¸çå³é®ç¹æ£æµé¨åç´§è·æåè¿çç®æ³ï¼åæ¬ Top-Down å Bottom-Up ä¸¤ç§æ¹æ³ï¼å¯ä»¥æ»¡è¶³ç¨æ·çä¸åéæ±ãåæ¶ï¼PaddleDetection æä¾éå¯¹ç§»å¨ç«¯è®¾å¤ä¼åçèªç å®æ¶å³é®ç¹æ£æµæ¨¡å PP-TinyPoseã

ä¼ éé¨ï¼PP-TinyPoseè¯´æã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

æ¨¡ååç§°	æ¨¡åç®ä»	COCOç²¾åº¦ï¼APï¼	éåº¦(FPS)	æ¨èé¨ç½²ç¡¬ä»¶	éç½®æä»¶	æ¨¡åä¸è½½
PP-TinyPose	è½»éçº§å³é®ç¹ç®æ³ è¾å¥å°ºå¯¸256x192	68.8	éªé¾865 åçº¿ç¨: 158.7 FPS	ç§»å¨ç«¯ãåµå¥å¼	é¾æ¥	ä¸è½½å°å

ä¼ éé¨ï¼å¨é¨é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
è¿å¨	å¥èº«	æä¾ä»æ¨¡åéåãæ°æ®åå¤ãæ¨¡åè®ç»ä¼åï¼å°åå¤çé»è¾åæ¨¡åé¨ç½²çå¨æµç¨å¯å¤ç¨æ¹æ¡ï¼ææè§£å³äºå¤æå¥èº«å¨ä½çé«æè¯å«ï¼æé AIèæå¥èº«æç»ï¼	åºäºPP-TinyPoseå¢å¼ºççæºè½å¥èº«å¨ä½è¯å«	ä¸è½½é¾æ¥

ðð»PP-Human å®æ¶è¡äººåæå·¥å·

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-Humanè¡äººåæå·¥å·ä½¿ç¨æåã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

ä»»å¡	T4 TensorRT FP16: éåº¦ï¼FPSï¼	æ¨èé¨ç½²ç¡¬ä»¶	æ¨¡åä¸è½½	æ¨¡åä½ç§¯
è¡äººæ£æµï¼é«ç²¾åº¦ï¼	39.8	æå¡å¨	ç®æ æ£æµ	182M
è¡äººè·è¸ªï¼é«ç²¾åº¦ï¼	31.4	æå¡å¨	å¤ç®æ è·è¸ª	182M
å±æ§è¯å«ï¼é«ç²¾åº¦ï¼	åäºº 117.6	æå¡å¨	ç®æ æ£æµ å±æ§è¯å«	ç®æ æ£æµï¼182M å±æ§è¯å«ï¼86M
æåè¯å«	åäºº 100	æå¡å¨	å¤ç®æ è·è¸ª å³é®ç¹æ£æµ åºäºå³é®ç¹è¡ä¸ºè¯å«	å¤ç®æ è·è¸ªï¼182M å³é®ç¹æ£æµï¼101M åºäºå³é®ç¹è¡ä¸ºè¯å«ï¼21.8M
é¯å¥è¯å«	31.4	æå¡å¨	å¤ç®æ è·è¸ª	182M
ææ¶è¯å«	50.8	æå¡å¨	è§é¢åç±»	90M
æ½çè¯å«	340.1	æå¡å¨	ç®æ æ£æµ åºäºäººä½idçç®æ æ£æµ	ç®æ æ£æµï¼182M åºäºäººä½idçç®æ æ£æµï¼27M
æçµè¯è¯å«	166.7	æå¡å¨	ç®æ æ£æµ åºäºäººä½idçå¾ååç±»	ç®æ æ£æµï¼182M åºäºäººä½idçå¾ååç±»ï¼45M

ä¼ éé¨ï¼å®æ´é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
æºè½å®é²	æåæ£æµ	é£æ¡¨è¡äººåæPP-Humanä¸æä¾çæåè¯å«ç®æ³ï¼éç¨äºå³é®ç¹+æ¶ç©ºå¾å·ç§¯ç½ç»çææ¯ï¼å¯¹æåå§¿å¿æ éå¶ãèæ¯ç¯å¢æ è¦æ±ã	åºäºPP-Human v2çæåæ£æµ	ä¸è½½é¾æ¥
æºè½å®é²	ææ¶è¯å«	æ¬é¡¹ç®åºäºPaddleVideoè§é¢å¼åå¥ä»¶è®ç»ææ¶è¯å«æ¨¡åï¼ç¶åå°è®ç»å¥½çæ¨¡åéæå°PaddleDetectionçPP-Humanä¸ï¼å©åè¡äººè¡ä¸ºåæã	åºäºPP-Humançææ¶è¯å«	ä¸è½½é¾æ¥
æºè½å®é²	æåæ£æµ	åºäºPP-Humanå®ææ¥å®¢åææ´ä½æµç¨ãä½¿ç¨PP-Humanå®ææ¥å®¢åæä¸éå¸¸å¸¸è§çåºæ¯ï¼ 1. æ¥å®¢å±æ§è¯å«(åéåè·¨å¢å¯è§åï¼ï¼2. æ¥å®¢è¡ä¸ºè¯å«ï¼æåè¯å«ï¼ã	åºäºPP-Humançæ¥å®¢åææ¡ä¾æç¨	ä¸è½½é¾æ¥

ðï¸PP-Vehicle å®æ¶è½¦è¾åæå·¥å·

ç®ä»(ç¹å»å±å¼)

ä¼ éé¨ï¼PP-Vehicleè½¦è¾åæå·¥å·æåã

é¢è®ç»æ¨¡å(ç¹å»å±å¼)

ä»»å¡	T4 TensorRT FP16: éåº¦(FPS)	æ¨èé¨ç½²ç¡¬ä»¶	æ¨¡åæ¹æ¡	æ¨¡åä½ç§¯
è½¦è¾æ£æµï¼é«ç²¾åº¦ï¼	38.9	æå¡å¨	ç®æ æ£æµ	182M
è½¦è¾è·è¸ªï¼é«ç²¾åº¦ï¼	25	æå¡å¨	å¤ç®æ è·è¸ª	182M
è½¦çè¯å«	213.7	æå¡å¨	è½¦çæ£æµ è½¦çè¯å«	è½¦çæ£æµï¼3.9M è½¦çåç¬¦è¯å«ï¼ 12M
è½¦è¾å±æ§	136.8	æå¡å¨	å±æ§è¯å«	7.2M

ä¼ éé¨ï¼å®æ´é¢è®ç»æ¨¡åã

äº§ä¸åºç¨ä»£ç ç¤ºä¾(ç¹å»å±å¼)

è¡ä¸	ç±»å«	äº®ç¹	ææ¡£è¯´æ	æ¨¡åä¸è½½
æºæ§äº¤é	äº¤éçæ§è½¦è¾åæ	æ¬é¡¹ç®åºäºPP-Vehicleæ¼ç¤ºæºæ§äº¤éä¸æåéçè½¦æµéçæ§ãè½¦è¾è¿åæ£æµä»¥åè½¦è¾ç»æåï¼è½¦çãè½¦åãé¢è²ï¼åæä¸å¤§åºæ¯ã	åºäºPP-Vehicleçäº¤éçæ§åæç³»ç»	ä¸è½½é¾æ¥

ð¡äº§ä¸å®è·µèä¾

ä¼ éé¨ï¼äº§ä¸å®è·µèä¾å®æ´åè¡¨

ðä¼ä¸åºç¨æ¡ä¾

ä¼ éé¨ï¼ä¼ä¸åºç¨æ¡ä¾å®æ´åè¡¨

ðè®¸å¯è¯ä¹¦

æ¬é¡¹ç®çåå¸åApache 2.0 licenseè®¸å¯è®¤è¯ã

ðå¼ç¨

@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}
}

Top Related Projects

Convert designs to code with AI

Introducing Visual Copilot: A new AI model to turn Figma designs to high quality code using your components.

Try Visual Copilot

Top Related Projects

Quick Overview

Pros

Cons

Code Examples

Getting Started

Competitor Comparisons

Pros of TensorFlow Models

Cons of TensorFlow Models

Code Comparison

Pros of Detectron2

Cons of Detectron2

Code Comparison

Pros of mmdetection

Cons of mmdetection

Code Comparison

Pros of YOLOv5

Cons of YOLOv5

Code Comparison

Pros of Mask_RCNN

Cons of Mask_RCNN

Code Comparison

Pros of darknet

Cons of darknet

Code Comparison

Convert designs to code with AI

README

ðç®å½

ðç®ä»

ð£ææ°è¿å±

ð«å¼æºç¤¾åº

â¨ä¸»è¦ç¹æ§

ð§©æ¨¡ååè®¾è®¡

ð±ä¸°å¯çæ¨¡ååº

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å ·

ð¡ðäº§ä¸çº§é¨ç½²å®è·µ

ð±å®è£

ð¥æç¨

ðFAQ

ð§©æ¨¡åç»ä»¶

ð±æ¨¡ååº

âï¸æ¨¡åæ§è½å¯¹æ¯

ð¥ï¸æå¡å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

âï¸ç§»å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å ·

ðPP-YOLOE é«ç²¾åº¦ç®æ æ£æµæ¨¡å

ðPP-YOLOE-R é«æ§è½æè½¬æ¡æ£æµæ¨¡å

ðPP-YOLOE-SOD é«ç²¾åº¦å°ç®æ æ£æµæ¨¡å

ð«PP-PicoDet è¶ è½»éå®æ¶ç®æ æ£æµæ¨¡å

ð¡PP-Tracking å®æ¶å¤ç®æ è·è¸ªç³»ç»

â·ï¸PP-TinyPose äººä½éª¨éª¼å ³é®ç¹è¯å«

ðð»PP-Human å®æ¶è¡äººåæå·¥å ·

ðï¸PP-Vehicle å®æ¶è½¦è¾åæå·¥å ·

ð¡äº§ä¸å®è·µèä¾

ðä¼ä¸åºç¨æ¡ä¾

ðè®¸å¯è¯ä¹¦

ðå¼ç¨

Top Related Projects

Convert designs to code with AI

ðç®å½

ðç®ä»

ð£ææ°è¿å±

ð«å¼æºç¤¾åº

â¨ä¸»è¦ç¹æ§

ð§©æ¨¡ååè®¾è®¡

ð±ä¸°å¯çæ¨¡ååº

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·

ð¡ðäº§ä¸çº§é¨ç½²å®è·µ

ð±å®è£

ð¥æç¨

ðFAQ

ð§©æ¨¡åç»ä»¶

ð±æ¨¡ååº

âï¸æ¨¡åæ§è½å¯¹æ¯

ð¥ï¸æå¡å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

âï¸ç§»å¨ç«¯æ¨¡åæ§è½å¯¹æ¯

ðï¸äº§ä¸ç¹è²æ¨¡å|äº§ä¸å·¥å·

ðPP-YOLOE é«ç²¾åº¦ç®æ æ£æµæ¨¡å

ðPP-YOLOE-R é«æ§è½æè½¬æ¡æ£æµæ¨¡å

ðPP-YOLOE-SOD é«ç²¾åº¦å°ç®æ æ£æµæ¨¡å

ð«PP-PicoDet è¶è½»éå®æ¶ç®æ æ£æµæ¨¡å

ð¡PP-Tracking å®æ¶å¤ç®æ è·è¸ªç³»ç»

â·ï¸PP-TinyPose äººä½éª¨éª¼å³é®ç¹è¯å«

ðð»PP-Human å®æ¶è¡äººåæå·¥å·

ðï¸PP-Vehicle å®æ¶è½¦è¾åæå·¥å·

ð¡äº§ä¸å®è·µèä¾

ðä¼ä¸åºç¨æ¡ä¾

ðè®¸å¯è¯ä¹¦

ðå¼ç¨