I have been working through my own version of detecting swimming pools.
I have use the same imagery (albeit, a subset) used by ESRI.
I think I have taken reasonable training sites. I have tried to be meticulous with selecting training sites. The ones I choose seem similar to those found in the ESRI examples.
I have used the same (and different parameters for training the model and detecting objects). I have taken the defaults, added values shown in ESRI examples, and tried some of my own options for parameters such as:
image tile, stride size
grid, scaling, ratio
padding, threshold, etc
epochs
The "learning curve" has shown concave to "jagged" (which I believe means, overfitted). Like I mentioned, I have tried this scenario numerous times with the different combinations of parameters for both training and detecting.
I am no where close to successfully detecting swimming pools. I mostly get features other than pools.
I am attaching my PPKX with the training and detected objects (and the image), about 23 Mb. I do have a model zip file, but it is approx. ~200 Mb, so it may be too large. I am attaching the EMD file and the PNG that shows the loss curve. This is one is jagged. 25 epochs.
The attachment has about 15 sample sites, but I have done this same process using a larger image and approx. 80 samples that I think are "good" samples of swimming pools...same kind of results.
If anyone can provide some suggestions or guidance, I would like to successfully attempt to find swimming pools....and then work with detecting other kinds of features.
I am using an XPS Dell business machine. Pro 2.5. I have the deep learning framework installed. No GPU or NVIDIA card. I can successfully run the full process, but the results are very bad.
I have viewed and use: Deep Learning: Detect swimming pools as a guide in addition to the ESRI help and other ArcGIS.learn notebooks and other information provided by ESRI.
Thank you for any insight or let me know if I should be contacting someone else at ESRI. I would also like to know if there is a process to go through to logically attempt different parameter values that can improve both training and detecting results.
Nate