One thing I always recommend when doing segmentation is first altering the color space; even utilizing a CNN (which should, when trained, essentially perform that step). This is especially true of pre-trained models where you don't know if it's been tuned for that.

Just in terms of edge detection, you'll see nearly a 10% improvement just from shifting the color space to LAB:

https://austingwalters.com/edge-detection-in-computer-vision...

What's even better about the color space conversion, is you can utilize it as a pre-processing step to dramatically reduce the number of edges to search:

https://austingwalters.com/chromatags/

In the case of a blue sea, you can shift to the LAB color space and probably only search the 'A' channel; the channel representing green to red. As the 'B' channel represents yellow to blue. Which is less likely to produce edges in the ocean.

This means you only process 1 channel, which dramatically speeds up most algorithms.

Just food for thought.

The related L-star c-star h-star works well too, and is easier to interpret than LAB.

These spaces benefit from being approximately perceptually linear, which also helps when any metric is computed on them.

Great meta article. Some other general advice concerning iterative development:

- Test your ideas on a smaller data sample at first to iterate and debug more rapidly.

- Caching your augmented data to disk may also improve iteration speed. Or it may not, depending on kernal disk cache vs disk bandwidth vs compute tradeoffs.

- Wrap your functions in timers so you can notice when a new processing stage is unnaturally slow, rather than having to back out later what happened.

- Make sure you're actually setup and using the GPU and not accidentally running on the CPU.