Mitsubishi Electric Innovation Center · Sep – Nov 2024
Optimizing YOLO detection models
- Role
- AI Intern — inference speed and accuracy across several object detection problems
- Dates
- Sep – Nov 2024 · Remote (Silicon Valley)
- Stack
- YOLOObject detection
- inference time
- −28%
- detection accuracy
- +18%
- real-time performance
- +22%
The problem
A detection model that clears an accuracy bar is not done — it also has to be fast enough for the setting it runs in. The work was to push YOLO-based detectors on both axes at once, and to do it across several different detection problems rather than one tuned showcase.
What I did
This was work inside a corporate innovation center, so I keep the detail here to what is on my CV; the specifics belong to the employer.
Results
- Inference time down 28%.
- Detection accuracy up 18%.
- Real-time performance up 22%.