search

developmentseed / label-maker

Data Preparation for Satellite Machine Learning
http://devseed.com/label-maker/
MIT License
461 stars 110 forks source link

Tensorflow Inference over satellite image tiles #126

Open Geo-Trackers opened 5 years ago

Geo-Trackers commented 5 years ago

Hello Everyone,

I want to detect brick-kilns in satellite imagery. I trained my model using Tensorflow object detection API. Now, I want to run inference over my satellite image tiles. I want to use gdaltools to generate image tiles and run inference over each tiles, and stich them back together for visualization over the large area of satellite imagery? Also, if my brick kiln falls in two image tiles, will the script tf_od_predict.py compute non-maximum suppression and get rid of low probability predictions, while i stich the images tiles back together. Thank you. I look forward to hear from you.

drewbo commented 5 years ago

Hi @Geo-Trackers. The tf_od_predict.py script runs on a per tile basis so it won't deduplicate predictions between two tiles. Hope that helps

Geo-Trackers commented 5 years ago

Does that mean I can run inference on per tile basis, and merge all tiles as a web map service to view my predictions with ease using the script?

Thank you for trying to clear the confusion. I really appreciate it.

Best Suman

Geoyi commented 5 years ago

You're right here, @Geo-Trackers. However, a large object may cross two tiles or more tiles so can write a utility function to get rid of bounding boxes that overlap by 75% threshold for instance. But I will recommend you to test with different threshold base on your case and object.