YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/
Features
Anchor-Free Detection Models - Provides an anchor-free YOLO architecture for real-time object detection with multi-backend inference support.
Real-Time Object Detection - Runs anchor-free object detection on images or video streams with higher accuracy and faster inference than prior YOLO versions.
Anchor-Free Detection Logic - Implements anchor-free detection logic that regresses object locations directly from center points without predefined anchor boxes.
Decoupled Detection Heads - Ships a decoupled detection head that separates classification and regression into distinct branches for better accuracy.
End-to-End Training Pipelines - Provides an end-to-end training pipeline from data loading to loss computation without post-processing dependencies.
Multi-Backend Inference Support - Supports running inference across ONNX, TensorRT, ncnn, OpenVINO, and MegEngine backends for deployment flexibility.
SimOTA Label Assignments - Assigns positive samples dynamically using a simplified Optimal Transport algorithm for improved training efficiency.
Production Inference Exports - Exports trained models to production-ready formats for deployment on diverse hardware platforms.
YOLO Object Detectors - Runs real-time object detection using a high-performance anchor-free YOLO architecture on images and video streams.
MegEngine Backend Integrations - Integrates MegEngine as the primary deep learning backend for automatic differentiation and model optimization.
Model Inference Deployment - Exports trained detection models to ONNX, TensorRT, ncnn, OpenVINO, or MegEngine for production inference.
Model Export Pipelines - Converts trained models to ONNX, TensorRT, ncnn, OpenVINO, and MegEngine formats for deployment across diverse hardware.
Multi-Format - Converts trained models to ONNX, TensorRT, ncnn, OpenVINO, and MegEngine formats for diverse hardware backends.
NMS-Free Inference Workflows - Eliminates non-maximum suppression during inference to reduce post-processing latency for real-time applications.
Edge Object Detection - Optimizes lightweight model variants for deployment on resource-constrained edge devices like mobile phones.
Inference Abstractions - Wraps model export and inference logic behind a unified API that abstracts hardware-specific optimizations.
Cross-Platform Deployments - Deploys object detection models across CPU, GPU, and edge devices using multiple inference backends.
Detection Model Exporters - Ships a dedicated export pipeline converting trained detection models to ONNX, TensorRT, ncnn, OpenVINO, and MegEngine formats.