AI for improving photogrametry

WebODM
1

Hello, seing the latest improvements in AI

Do you think you guys could incorporate AI for improving photogrametry results?

Thanks

2

I’ve read about that, very interesting.

But I don’t think it’s so easy to implement.

3

I’m cautious here.

AI at the core is statistical inference based on patterns that are seen by some code - so everything generative is more or less something already seen recycled (in a way as researchers we do the same when we make empirical models of stuff based on observations. We make decent guesses, and know we don’t represent reality a whole lot of the time)

Filling in detail based on inference is fine if the end point is a prettier model for real estate or game environments.

It is less fine if we’re seeing new things - eg research data, something which doesn’t fit the mold. Or we need accurate representations of real things - inspection, surveying…

Another related question is: what is better way to spend energy - levelling up flight and data productions skills, or getting someone to burn a bunch of compute time training models?

My money is on the former - if we want better models, level ourselves up and fly better :wink:

4 Likes
4

You elucidated my feelings on this wonderfully, and far kinder than I was able to find words for :rofl:

There are applications, for sure, but there need to be serious guard-rails in place.

1 Like
5

i see DJI terra can real time build 3D map, do you think they use AI for improve performance of photogrametry?

1 Like
6

I have not looked at the terra but I think they are just doing some SLAM here.

2 Likes
7

I’ve thought about classifying point cloud data. This would enable WebODM to both classify photogrammetry and LiDAR data in one platform. Would like to meet folks who want to brainstorm on this – please ping me.

2 Likes
8

Hi Tariqislam,
I am researching image processing for drones at a startup company, I try to understand algorithm inside opendronemap, can i ping you and ask some question?
cheers,
TraDinh,

1 Like
9

Hi,

Sorry I am new to image processing, and I don’t know how WebODM processing pipeline works either. I’m also trying to learn, so when you find the resource please ping me as well!

Tariq

1 Like
10

I personally think that utilizing AI with drone data is going to be a game changer across the board. But I think at this the time AI component should be more focused on the dataset post processing. IE - Distinguishing buildings, landscapes, trees etc. Things of that nature. Eventually implementing it into the processing side of things as well. It’s coming but I don’t think we’re there yet.

2 Likes
11

Check out Piero’s OpenPointClass:

4 Likes
12

Saijin,
I looked at this and am super intrigued. How does one go about creating labeled data to input into the training model? Is there a video showing the software being used? Super interested in this.

2 Likes
13

I’d recommend using something like CloudCompare or QGIS for manual classification of PointCloud data, then following along with this:

We’re working on a process that will make it easier for folks to contribute manually classified/verified datasets to help bolster our model.

1 Like
14

Awesome! I’ll have to look into it a little more but this is going to be great once we’re able to put in a large amount of classified datasets!

2 Likes
15

This looks awesome. Would love to know when/how I can help.

1 Like
16

Me too!

I don’t understand our process well enough yet, and I’ve not worked through manually to try and pick out any gotchas.

Once I get my head around it a bit better, we’ll have some (hopefully very friendly) workflows for folks to follow.

2 Likes
17

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.