| Property | Value |
|---|---|
| Author | Ao Wang et al. |
| Paper | arXiv:2503.07465 |
| Model Sizes | S, M, L variants |
| Parameters | 12M-50M |
What is YOLOE?
YOLOE (YOLO Eye) is a groundbreaking unified object detection and segmentation model that introduces real-time "seeing anything" capabilities. It uniquely combines efficiency with versatility by supporting multiple prompt mechanisms - text, visual, and prompt-free paradigms - all within a single model architecture.
Implementation Details
The model implements three key innovations: Re-parameterizable Region-Text Alignment (RepRTA) for text prompts, Semantic-Activated Visual Prompt Encoder (SAVPE) for visual prompts, and Lazy Region-Prompt Contrast (LRPC) for prompt-free scenarios. These components enable state-of-the-art performance while maintaining high inference efficiency.
- Multiple model variants (v8-S/M/L and 11-S/M/L) with different parameter sizes
- Achieves up to 305.8 FPS on T4 GPU
- Supports both detection and segmentation tasks
- Zero-shot capabilities on LVIS dataset
Core Capabilities
- Real-time object detection and segmentation
- Multi-prompt support (text, visual, prompt-free)
- Efficient re-parameterization for transfer learning
- Superior performance compared to YOLO-Worldv2
- CoreML and TensorRT deployment support
Frequently Asked Questions
Q: What makes this model unique?
YOLOE's uniqueness lies in its ability to handle multiple prompt types within a single efficient architecture, while achieving real-time performance. It offers 3× less training cost and 1.4× inference speedup compared to similar models.
Q: What are the recommended use cases?
The model is ideal for real-time object detection and segmentation applications, especially in scenarios requiring flexible object recognition without predefined categories. It's particularly suitable for deployment on both GPU (T4) and mobile devices (iPhone).