PaddleDetection v2.7.0

PaddleDetection flying paddle target detection development kit is designed to help developers complete the entire development process of detection model construction, training, optimization and deployment faster and better.

PaddleDetection modularly implements a variety of mainstream target detection algorithms, provides rich data enhancement strategies, network module components (such as backbone networks), loss functions, etc., and integrates model compression and cross-platform high-performance deployment capabilities.

After long-term industrial practice and polishing, PaddleDetection has a smooth and excellent user experience, and is widely used by developers in more than ten industries such as industrial quality inspection, remote sensing image detection, unmanned inspection, new retail, Internet, and scientific research.

characteristic

Rich models: including 100+ pre-trained models such as target detection, instance segmentation, face detection, etc., covering a variety of global competition championship solutions

Simple to use: modular design, decoupling various network components, developers can easily build and try various detection models and optimization strategies, and quickly obtain high-performance, customized algorithms.

End-to-end connectivity: End-to-end connectivity from data enhancement, networking, training, compression, and deployment, and fully supports cloud/edge multi-architecture and multi-device deployment.

High performance: Based on the high-performance core of the flying paddle, the model training speed and memory usage are obvious. Supports FP16 training and multi-machine training.

PaddleDetection v2.3.0 change log

Model richness

Released Transformer detection models: DETR, Deformable DETR, Sparse RCNN

Added new Dark model for key point detection and released Dark HRNet model

Released MPII dataset HRNet key point detection model

Publish head and vehicle tracking vertical models

Model optimization

The rotating frame detection model S2ANet releases the Align Conv optimization model, and the DOTA data set mAP is optimized to 74.0

Predictive deployment

Mainstream models support batch size>1 prediction deployment, including YOLOv3, PP-YOLO, Faster RCNN, SSD, TTFNet, FCOS

Added support for Python-side prediction deployment of multi-target tracking models (JDE, FairMot, DeepSort), and supports TensorRT prediction

Added multi-target tracking model FairMot joint key point detection model deployment Python side prediction deployment support

New key point detection model combined with PP-YOLO prediction deployment support

document

New TensorRT instructions added to Windows Predictive Deployment documentation

FAQ document update released

Bug fixes

Fix the convergence problem of PP-YOLO series model training

Fix the problem of unlabeled data training when batch size>1