THU-MIG/yolov10

Features

• NMS-Free Object Detection - Provides an end-to-end detection architecture that eliminates non-maximum suppression to reduce inference latency.
• YOLO Object Detectors - Implements an end-to-end YOLO object detection architecture optimized for real-time identification and location of objects.
• Computer Vision Libraries - Ships a comprehensive computer vision library for training, validating, and exporting real-time detection models.
• Real-Time Object Detection - Enables real-time object detection in images and video streams with minimized post-processing latency.
• Detection Model Training - Provides tools to train object detection models on custom datasets using specified epochs and batch sizes.
• NMS-Free Object Detectors - Functions as an end-to-end object detector that eliminates NMS to maximize inference speed.
• Inference Latency Optimizers - Reduces computational time and overhead to optimize inference latency for real-time visual processing.
• Edge AI Model Deployment - Optimizes and deploys machine learning models for efficient execution on local and edge devices.
• Edge Hardware Optimizations - Provides edge hardware optimizations including quantization to ensure high-speed deployment on constrained devices.
• Custom Vision Training - Supports custom vision training by optimizing neural networks to recognize specific object classes.
• Neural Network Architectures - Utilizes a specialized neural network architecture designed to minimize inference latency through a reduced backbone.
• PyTorch Computer Vision Pipelines - Ships a complete PyTorch-based computer vision pipeline for training and optimizing object detection models.
• PyTorch Training Frameworks - Integrates PyTorch training frameworks to optimize model weights on custom datasets across scalable sizes.
• Efficient Neural Architectures - Employs a lightweight neural network structure as a high-performance backbone to reduce computational overhead.
• Real-Time Model Inference on Frames - Provides a framework designed for high-speed ML inference on live video frames for real-time detection.
• Spatial Attention Weighting - Integrates spatial attention weighting to emphasize relevant features and improve accuracy in complex scenes.
• NMS-Free Inference Workflows - Implements an inference workflow that predicts a single box per object, bypassing non-maximum suppression.
• Multi-Scale Feature Aggregation - Merges feature maps across diverse spatial scales to accurately identify objects of varying sizes.
• Label Assignment Strategies - Implements dual-label assignment strategies using hard and soft labels to optimize training convergence and detection precision.
• Detection Model Validation - Includes capabilities to evaluate detection accuracy using metrics like mean average precision against benchmark datasets.
• Precision Quantization - Provides precision quantization of weights to enable high-speed execution on edge hardware.
• Weight Quantization - Organizes model weights using quantization techniques to optimize execution for resource-constrained edge hardware.
• Model Weight Conversions - Provides utilities for converting model weights into optimized formats compatible with various hardware accelerators.
• Computer Vision - Real-time object detection with NMS-free training.
• Foundational AI References - Innovations in computer vision and object detection.
• Object Detection - Listed in the “Object Detection” section of the The Incredible Pytorch awesome list.

Star history