Things I'd do:

• Much cleaner UI for all of the model metrics (currently can't compare models very well at all)
• More info on memory useage and inference speeds of different models in different formats (on different hardware).
• Much more modern and more searchable documentation, actually annoate the code and integrate code with docs.
• More demos
• Integrate directly with more modern Computer Vision tooling (e.g. voxel51, CVAT and supervision)