R

Rtdetr R18vd Coco O365

Developed by PekingU
The first real-time end-to-end object detector, achieving efficient NMS-free detection through hybrid encoder and query selection mechanism
Downloads 952
Release Time : 5/21/2024

Model Overview

RT-DETR is a Transformer-based real-time object detection model that significantly improves detection speed while maintaining accuracy by eliminating the Non-Maximum Suppression (NMS) process. Supports flexible speed adjustment by varying decoder layers.

Model Features

NMS-free design
Eliminates Non-Maximum Suppression in traditional object detection through end-to-end Transformer architecture, reducing computational overhead
Hybrid encoder
Combines intra-scale interaction with attention (AIFI) and CNN-based cross-scale fusion (CCFF) for efficient multi-scale feature processing
Dynamic speed adjustment
Supports varying inference speeds by adjusting decoder layers without retraining, adaptable to multiple scenarios
Query selection mechanism
Provides high-quality initial queries for the decoder through minimum uncertainty query selection, improving detection accuracy

Model Capabilities

Real-time object detection
Multi-scale object recognition
End-to-end bounding box prediction

Use Cases

Smart surveillance
Public space crowd analysis
Real-time detection of crowd density and movement trajectories in airports, stations, etc.
108FPS@53.1AP (R50 model)
Autonomous driving
Road object detection
Identifying vehicles, pedestrians, traffic signs and other road elements
Supports different speed level configurations (74-217FPS)
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase