I am trying to run arcgis.learn.detect_object() function on an Image Server (ArcGIS Server 10.8.1) with 2 VM (2 GPUs Nvidia Tesla T4, 16GB RAM each one). I use the “Building Footprint Extraction – USA” as model and an image (5KmX2km, 0.1m ps) to test the site.
Sometimes the code is running and the results are similar to ArcGIS Pro Image Analyst, but sometimes it crashes and it appears one of the next two messages.
When I try to test the code with a bigger image (e.g., a mosaic of 5-6 images as the above) it crushes all the time.
Finally, I’d like to ask if you have any suggestions about improving the performance and optimizing the parameters for better results. As you can see on the script, I have padding:32, threshold:0.9 and batch_size: 8. For a single image, it takes a few minutes with 2 GPUs (It takes about 2 hours with one GPU), but the results need improvement.