dog-qiuqiu/FastestDet

:zap: A newly designed ultra lightweight anchor free target detection algorithm， weight only 250K parameters， reduces the time consumption by 10% compared with yolo-fastest, and the post-processing is simpler

What it solves

FastestDet is an ultra-lightweight, real-time object detection algorithm designed for high performance on resource-constrained devices, such as ARM CPUs and NPUs. It aims to provide a faster and simpler alternative to previous lightweight detectors like Yolo-fastest, improving both detection accuracy (mAP) and inference speed.

How it works

The project implements several architectural optimizations to achieve its speed and efficiency:

  • Anchor-Free Design: Removes the need for predefined anchor boxes, simplifying the detection process.
  • Single Scale Detector Head: Reduces computational overhead by using a single scale for detection.
  • Cross Grid Multiple Candidate Targets: Enhances target localization.
  • Dynamic Positive and Negative Sample Allocation: Optimizes how the model learns to distinguish between objects and background.
  • Loss Optimization: Uses an IOU-aware approach based on smooth L1 to increase precision.

Who it’s for

This project is ideal for developers and researchers working on embedded systems, mobile devices, or any environment where CPU/NPU power is limited and real-time object detection is required.

Highlights

  • Ultra-Lightweight: Features a very small parameter count (0.24M).
  • High Efficiency: Optimized for deployment on platforms like Radxa Rock3A (RK3568), Qualcomm Snapdragon 835, and Intel i7-8700.
  • Multi-Framework Support: Supports exporting to ONNX, TorchScript, and deployment via NCNN and ONNX-Runtime.
  • Improved Performance: Outperforms Yolo-fastestv2 in both mAP (25.3% vs 24.1%) and run time on ARM CPUs.