iterative search is an important method to approach to the optimal parameter of a system. It works even the system is a black box whist providing a output response only. Here I list my thoughts on adopting iterative search method.

  • loss function is critical, which in most case requires expertise experience to decide.
  • ground truth: sometimes ground truth can be obtained via handcraft ways, which greatly helps.
  • parameter space: which is usually big. However, there are some skills, eg.,  checking the boundary parameters, focusing on some core parameters, provide insight and reveal the core problem.
  • parameter attributes: boundaries(min, max), type(float, int, str), static/variable/watching, name, etc.
  • initial value and re-start.

Note:

I adopt this way to find the optimal camera ISP parameters. photoshop is used to generated groundtruth image with RAW file. However, I tried quite a few loss function, but a better one is still in need. I may try S3 if possible. And suggestion is highly appreciated. <shock.jiang@gmail.com>

 

… to be continued …