Discussion

1. Relevance

There is an increasing scope for applying computer vision techniques for intelligent camera applications on smartphones. Many such applications will need to run in real-time to ensure a good user experience. With this in mind, our efforts to measure the performance of one common technique (object detection) can be helpful when thinking of developing a camera application using models like SSD-MobileNet and YOLO, as we can estimate run-time performance better.

Of course, we can extend this work directly to a "real" smartphone application as well.

2. Problems

(Some) Problems in this work:

• performance of YOLO is quite poor compared to SSD-MobileNet (both accuracy and speed); should investigate further
• models with quantized weights didn't work out; would be a useful data point if they can be made to work

3. Possible future work

• A more thorough profile of the time taken by various components of the application
  • image acquisition
  • pre-processing
  • network inference
  • output post-processing (non-maximal suppression)
• Optimizing the application based on said profile

• Re-training YOLO with computational tweaks (depthwise convolution, etc.)
  • Maybe with YUV input?

• Extending to PyTorch, Darknet, etc. using Native Design Kit