Select to view content in your preferred language

Object detection using deep learning doesn't work well

945
2
10-29-2023 12:01 AM
NIKOLAYGK
New Contributor

Hello. I am using ArcGIS Pro 3.1 to detect for objects in an image. I'm doing everything right, but the result is unsatisfactory. I've tried changing the number of max epochs, the type of model (SSD and RCNN), tile sizes, the results don't get any better. Below are screenshots of the settings with which I managed to achieve the best results (but they still remain very bad).
Maybe the training sample was compiled by me incorrectly? I really hope that someone will help me solve this problem.    

1.jpg

2.jpg

3.jpg

4.jpg

5.jpg

6.jpg

7.jpg

8.jpg

9.jpg

10.jpg

11.jpg

12.jpg

   

2 Replies
David_McRitchie
Esri Contributor

Hey,

Are you still facing issues with this? I was just taking a look at unanswered threads and came across this.

If you're hitting issues then it would be good to know how you have done your training data and what specific features you are trying to detect.

Many thanks,

David

Esri UK -Technical Support Analyst
0 Kudos
ShivaniPathak
Esri Contributor

Hey,

The labels which you created are masks so from my understanding you want to get masks of heracluemm, for getting results in the form of masks instead of a bounding box you can train either a pixel classification model (i.e. UnetClassifier, DeepLab, etc) or instance segmentation model (i.e. MaskRCNN). For training a pixel classification model you will need to select Metadata Format parameter as Classified Tiles in Export Training Data For Deep Learning tool and inference using the Classify Pixels Using Deep Learning tool.

For training an instance segmentation model you will need to select Metadata Format parameter as RCNN Masks in Export Training Data For Deep Learning tool and inference using the Detect Objects Using Deep Learning tool.