RangiLyu/nanodet

Features

• Lightweight Anchor-Free Detectors - Ships a 980KB anchor-free detection model achieving 97 FPS on mobile devices for real-time edge inference.
• Lightweight Anchor-Free Detections - Provides a lightweight anchor-free detection model optimized for speed and accuracy on resource-constrained devices.
• Anchor-Free Detection Logic - Implements anchor-free detection logic that predicts bounding boxes directly from feature map points.
• GhostNet Backbones - Provides a GhostNet backbone that reduces parameters and FLOPs through cheap linear operations.
• Real-Time Object Detection - Achieves 97 FPS real-time object detection on mobile devices for live video stream analysis.
• Depthwise Separable Convolutions - Uses depthwise separable convolutions to shrink model size and accelerate inference.
• Detection Model Training - Provides a training pipeline for custom object detection models with configurable backbones and data augmentation.
• Light-Weight Detection Necks - Ships a light-weight detection neck that aggregates multi-scale features with minimal overhead.
• Edge AI Model Deployment - Exports and deploys detection models to edge hardware using ONNX, ncnn, MNN, or OpenVINO.
• ONNX Model Exporters - Exports object detection models to ONNX format for cross-platform deployment.
• Mobile-Optimized Neural Networks - Designs neural network architectures for high-speed object detection on ARM CPUs and mobile platforms.
• PyTorch-to-ONNX Pipelines - Converts trained PyTorch models to ONNX format for deployment across multiple inference engines.
• Python API - Provides a Python API for running object detection inference on images, videos, or webcam streams.
• Real-Time Model Inference on Frames - Performs object detection on live video streams at high frame rates on low-power hardware.
• Edge Deployments - Deploys trained detection models to edge devices for real-time inference using ONNX and ncnn/MNN/OpenVINO.
• ARM-Optimized Detectors - Delivers a 1.8MB model optimized for ARM CPUs achieving 97 FPS on mobile devices.
• Multi-Backend Inference Support - Supports running inference across multiple backends including ncnn, MNN, OpenVINO, and web browsers.
• Assign Guidance Modules - Ships an Assign Guidance Module that improves training target assignment using classification and regression quality.
• Dynamic Soft Label Assignments - Implements dynamic soft label assignment to balance positive and negative samples during training.
• CNN - Listed in the “CNN” section of the Ailia Models awesome list.

Star history