Olympus
Olympus is a universal task router designed for computer vision tasks, recognized as a highlight at CVPR 2025. Developed by researchers including Yuanze Lin and Philip H. S. Torr, this software serves as an official implementation that directs various vision-related operations, leveraging a specialized routing mechanism to handle complex instructions. The system integrates with task-specific models to enhance efficiency and adaptability across diverse applications. Key features include a training and inference framework, pre-trained models available on Hugging Face, and proprietary datasets covering twenty distinct tasks. The project provides comprehensive resources such as benchmark datasets, fine-tuning instructions, and support data including chain-of-action examples. Users can install the environment via Conda and Python, download the model checkpoints and dataset files through provided scripts, and access detailed documentation via the project webpage. Olympus aims to streamline the deployment of compute