How to Speed Up Deep Learning Capability for Segmentation and Object Detection in ArcGIS Online?

[Joy_Chan]

My client from University said he was excited about our deep learning capability for segmentation and object detection, making use of pre-trained models now available on ArcGIS Online.

He first tried to work on a satellite image with 2000-3000 houses, trying to identify the footprint of the houses, and it took him 30 mins till he stopped, as it took too long. He then picked a much smaller area with about 100 houses, this time it took him 15 minutes to run. He thinks it’s still not acceptable with this speed as he wants to introduce this tool to his students during his class. Is this speed affected by the ArcGIS Online Cloud server computation power? How could it be enhanced?

[David_McRitchie]

Hey Joy,

It will really depend how they are running this. Generally, for best performance I recommend doing Deep Learning within ArcGIS Pro with a machine that has a high Cuda Compute capability.

Using a small sample area is recommended, as it is worth testing different variables to the Deep Learning model to see if performance can be improved before using the model on your full dataset. In my experience, increasing the Batch_Size will make a big difference to the processing time.

Ultimately however performance is subjective and I would say 15 minutes is quite quick for a deep learning workflow so it is key to experiment with different configurations on the smallest sample area and find what works best for your workflow.

Hope that helps,

David