support YOLO-World [Ultralytics version]

[ZhengQuancheng]

Description

This PR adds support for YOLO-World models to DeepStream.

YOLO-World is an open-vocabulary detection model that can detect objects based on text descriptions. Unlike traditional YOLO models, YOLO-World uses CLIP text embeddings to enable flexible class definition without retraining.

Features

• Support for all YOLO-World model variants (yolov8s/m/l/x-worldv2)
• Custom class support via --custom-classes parameter (e.g., --custom-classes "person, car, dog")
• Default COCO 80 classes when no custom classes specified
• CLIP text embeddings are pre-computed and baked into the ONNX model for efficient inference

Files Added

• utils/export_yoloworld.py - Export script for YOLO-World models
• docs/YOLOWorld.md - Documentation for YOLO-World usage
• config_infer_primary_yoloworld.txt - DeepStream inference configuration
• README.md - Update content related to YOLO-World

Usage

# Export with default COCO 80 classes
python3 export_yoloworld.py -w yolov8s-worldv2.pt --dynamic --simplify

# Export with custom classes
python3 export_yoloworld.py -w yolov8s-worldv2.pt --custom-classes "person, car, dog" --dynamic --simplify

Testing

• Tested on DeepStream 8.0 with CUDA 12.8
• Successfully exported ONNX model and generated TensorRT engine
• Verified inference pipeline with sample video