YOLOS Small

YOLOS-small is an object detection model that applies a Vision Transformer (ViT) backbone to the object detection task, trained with the DETR bipartite matching loss and fine-tuned on COCO 2017.

Architecture and training

The model uses a vanilla ViT architecture with minimal spatial priors. It is pre-trained on ImageNet-1k for 200 epochs and then fine-tuned on COCO 2017 object detection (118k training images) for 150 epochs. Detection is performed by adding 100 learnable detection tokens to the input sequence; the model predicts class and bounding box for each token. The Hungarian algorithm matches predictions to ground truth, and the loss combines cross-entropy for classes with L1 and generalized IoU for boxes.

Performance

On COCO 2017 validation, YOLOS-small achieves an average precision (AP) of 36.1. For comparison, the larger YOLOS-base variant reaches 42.0 AP, matching more complex frameworks such as Faster R-CNN and DETR. The model contains 30.7 million parameters and is released under the Apache-2.0 license.

References

• Paper: You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection (Fang et al., NeurIPS 2021)
• Code and pre-trained models: github.com/hustvl/YOLOS
• Hugging Face model page: hustvl/yolos-small