YOLOS

Why we included this project

YOLOS is a compact research codebase that shows how far a plain Vision Transformer can go on object detection. Instead of stacking convolutional feature pyramids and task-specific heads, it treats detection as a sequence problem: image patches go through an unmodified ViT backbone, and the model learns to emit detection tokens. The repo includes pretrained weights and an attention visualization notebook, so it is straightforward to reproduce the NeurIPS 2021 paper's results or compare a pure transformer detector against the hybrid designs that followed. Since the model is also integrated into Hugging Face Transformers, you can load and run it without compiling the source yourself. For anyone curious about whether transformer-only architectures can replace the usual detection backbones, this is a tidy baseline to study.