Future of Dashcam-Lok

What are some problems the follow-up project to Dashcam-Lok could focus on?

Assuming that the current phase of Dashcam-Lok ends with delivering:

• 50k hours of data (RGB, with metadata and possibly unanonymized)
• Model trained for 'person' deployed on the edge-device
• Model trained for the other 'n' classes deployed on the workstation

And probably delivering:

• The trained model(s) and weight(s)
• The labeled data used for train/val/test
• Statistics - number of classes and their frequencies

What we don't have:

• Annotations for the collected data
• Depth data - relative depth and absolute depth data

Ideas
Data - Focusing on scaling up data collection
AI - Focusing on using the collected data for further research