#How does angle effect the change detection AI model

23 messages · Page 1 of 1 (latest)

hidden abyss
#

I'm starting to work on a change detection model that will detect change based on images (of a city for example) taken by a drone and I was wondering if the angle that the picture was taken with, like if i would like to compare between two pictures but each of them is taken with different angle.
would that effect the efficiency of the model and if it does how can fix it?

jolly oracle
#

Sorry I don't have the solution but that sounds like a very tricky and worthwhile problem to solve! The AI would need to understand that objects are 3D I guess. You could pre train it with 3D rendered objects, give it pairs of screenshots of those objects with each pair classed as 'same' or 'different' maybe? I'm just an undergrad, but another very naive approach could be to run object detection on the image, and use that to establish change (but then, say you detect a dog, you wouldn't know if it's a different dog or not... Unless you require it to give a description?). Very interesting, do update if you find something!!

#

Of course you could always just feed it a bunch of images of cities from different angles to train it what 'same' and 'different' are, but then you're less sure of what it will learn and whether it will perform well in testing

cunning gale
#

a value indicating the amount of change?

#

specific localized changes?

hidden abyss
cunning gale
#

What is your model?

#

I'm assuming from the wording of the question that you already have a strategy

hidden abyss
#

I didn't start yet but I'm planning to use Unet

cunning gale
#

How are you going to use U-Net

#

You have two different image inputs

#

Also, U-Net outputs an image while you said that you want a value indicating the amount of change

#

I would recommend Siamese network but I'm not 100% sure if it would cope well with the different angles

hidden abyss
#

but i dont think that the model is the problem here, i think that the problem is with the provided data or how I'll provide the data to the model

#

like isn't there a way that i can tell the model that those two images that have different angels are the same?

cunning gale
#

This is really dependent on what you want to do with your outputs

#

If you are intending to make a Siamese network, then contrastive learning with triplets is probably your best bet

#

But U-Net cannot be trained contrastively (as far as I know)

cunning gale
hidden abyss
#

can I add you so I can ask you more questions if had any? since I need to dig more on the subject