D

Dab Detr Resnet 50 Dc5 Pat3

Developed by IDEA-Research
DAB-DETR is an improved DETR model that significantly enhances object detection performance and training efficiency by using dynamic anchor boxes as queries.
Downloads 31
Release Time : 9/10/2024

Model Overview

This model employs dynamic anchor boxes as queries for the Transformer decoder, updating box coordinates layer by layer, effectively addressing the slow convergence issue in traditional DETR training and demonstrating excellent performance on the COCO benchmark.

Model Features

Dynamic Anchor Box Queries
Directly uses box coordinates as queries in the Transformer decoder and dynamically updates them layer by layer, significantly improving training efficiency.
Soft ROI Pooling
Queries can perform soft ROI pooling layer by layer in a cascaded manner, optimizing the feature extraction process.
Efficient Training
Compared to traditional DETR, training converges faster, achieving optimal performance in just 50 epochs.

Model Capabilities

Image Object Detection
Multi-object Recognition
Bounding Box Prediction

Use Cases

Computer Vision
General Object Detection
Detects and localizes multiple objects in complex scenes
Achieves 45.7% AP on the COCO dataset
Intelligent Surveillance
Real-time detection of multiple targets in surveillance videos
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase