Questions About Object Detection with Deep Learning in ArcGIS Online

Question

I’m using ArcGIS Notebooks in ArcGIS Online with GPU support to train a deep learning model (MaskRCNN) and publish the package (.dlpk). The training and publishing work well, but object detection with DetectObjectsUsingDeepLearning in ArcPy takes about 15 hours, even with GPU.

Questions:

* Web Applications: Is it possible to use the deep learning package in web applications for real-time object detection?
* Optimization: How can I reduce the detection time? Is it normal for it to take this long even with GPU?
* Recommended Workflow: What is the most efficient workflow for object detection with ArcGIS Online?

Thank you for your assistance.

Best regards,

PriyankaTuteja · Answer

Hi Leonel,

The DetectObjectsUsingDeepLearning in ArcPy job could be run faster by increasing the batch size to ensure that the tool consumes your machines' GPU at full capacity. DLPK is recognized as a deep learning package that can be consumed by tools and our api. The use of dlpks is not supported outside of Esri's software.

LeonelRivero_79 · Answer

Hi David,

Thanks for your response.

Regarding the use of ArcPy, I’m curious if it’s the only method available for object detection once the model is published as a deep learning package (DLPK) in ArcGIS Online. Since the DLPK is recognized as an item in ArcGIS Online, does this indicate that there are other Python modules or methods I can use for detection, besides arcpy.DetectObjectsUsingDeepLearning? Or is the DLPK primarily intended as a stored file with limited usage outside of ArcPy?

I appreciate your insights.

Best regards,
Leonel

D_McRitchie · Answer

Hey Leonel,

These are very good questions. In regard to optimisation, performance with Deep Learning can be a difficult topic so it would be good to have further information on size and resolution of the imagery, as well as clipping of this.

Generally with any Deep Learning workflow I would take the smallest possible sample, with a few dozen training samples to gauge roughly how long a process might take,and to provide a testing environment where changes can be tested and quickly validated if they will improve detections and performance.

For your other questions it would be good if we can get some ideas on what objects you are trying to detect. I recommend starting with the Living Atlas as there are a wide range of pre-trained models that can be used and this may save a lot of time training your own model.

Hope that helps,

David