| Property | Value |
|---|---|
| Parameter Count | 30.7M |
| License | Apache 2.0 |
| Paper | View Paper |
| Performance | 36.1 AP on COCO |
What is yolos-small?
YOLOS-small is a compact Vision Transformer (ViT) model designed specifically for object detection tasks. Developed by hustvl, it represents a simplified approach to transformer-based object detection, achieving impressive results while maintaining a relatively small parameter count of 30.7M.
Implementation Details
The model employs a bipartite matching loss system and processes images through a transformer architecture. It handles 100 object queries simultaneously and uses the Hungarian matching algorithm to optimize object detection. The model was pre-trained on ImageNet-1k for 200 epochs and fine-tuned on COCO 2017 for 150 epochs.
- Utilizes PyTorch framework for implementation
- Supports F32 tensor operations
- Implements DETR-style loss function
- Combines L1 and generalized IoU loss for bounding boxes
Core Capabilities
- Object detection with state-of-the-art accuracy
- Processing of multiple object queries simultaneously
- Efficient feature extraction from images
- Real-time bounding box prediction
- COCO class classification
Frequently Asked Questions
Q: What makes this model unique?
YOLOS-small stands out for its simplicity and efficiency, achieving 36.1 AP on COCO validation while using a pure transformer-based architecture, eliminating the need for complex detection frameworks like Faster R-CNN.
Q: What are the recommended use cases?
The model is ideal for object detection tasks in real-world scenarios, particularly when working with the COCO dataset's object classes. It's especially suitable for applications requiring a good balance between model size and performance.