Fine-Tune deep learning detect objects - swimming pool

[Letíciad_Agosto_Miguel_Fonseca]

Hello everyone, 

I’m currently working on fine-tuning the pool detection model for use in southern Brazil. I followed the steps outlined in this post: https://doc.arcgis.com/en/pretrained-models/latest/imagery/finetuning-the-pool-detection-usa.htm

Raster information:

Training Samples:

Parameters:

Image training sample (0.3m):

However, the results I'm getting seem a bit unusual:

Does anyone have any suggestions on how I can improve these results?

Thanks in advance!

[PavanYadav]

Hi @Letíciad_Agosto_Miguel_Fonseca it does look concerning. The model is trained with 8-bit, 3-band high resolution (5-30 centimeters) imagery. What kind of imagery did you use to create the training samples. Can you share a screen capture of its stats.txt file?.it should look something like this example 

[Letíciad_Agosto_Miguel_Fonseca]

Hello @PavanYadav , first of all, thank you for your attention. Indeed, I noticed some inconsistencies.

Here is the stats.txt from my training samples:

The .TIFF image I exported from the .ECW file is showing 4 bands, although there are only 3.

I re-exported it using the Raster to Other Format geoprocessing tool and now it is right:

Now, the statistics for my training samples (62 swimming pools) are as follows:

However, I believe I need more training data and better resources to fine-tune the model. When I tried running the tool, my ArcGIS Pro crashed, and it's not feasible to use 100 epochs and a batch size of 64.

I’ll try running it on another server and see if I can successfully fine-tune the model.

If you have any more suggestions on how to improve, please let me know.

[PavanYadav]

@Letíciad_Agosto_Miguel_FonsecaSorry for the late response. ArcGIS Pro shouldn't crash even if the batch size exceeds GPU memory. Can you try using a smaller batch size, such as 2 or 4? More training samples should also help. Another thing to try is building a new model solely with your training data (don't use the pre-trained model as a starting point) to see if you get any meaningful results.

Cheers!