We released our new experimental model for 10cm images — the ones captured by drones or aerial sensors from above.
The new model is much more capable of detecting small buildings in the rural areas but at the same time it’s trained to skip non-capital structures like “greenhouses”.
What is the difference between these two maps below generated with Mapflow.ai? At this scale it’s hard to say. However, the image on the left is processed with the model trained on satellite imagery which labeled out 1994 features. While the image on the right is processed using the new Aerial/UAV model that labeled out 4186 features (x2.1 times more).
Getting a closer look — that what makes the difference. More features and more details.
The new model can better “see” the constructions partially hidden by the trees in the forestry areas — like this one, somewhere in Argentina.
If not simplified it looks like this:
Giving more examples of model testing on Mapflow.ai one can see for himself — there is no evidence for the best performance all over the world. The context matters. If the performance is poor, we have to extend the training dataset to tune the model up to the level which is…
Let’s test UAV model “in the wild” :-)
Alexander, our CV engineer, is on the Far East now, travelling across Kamchatka. Good for him!
I met a man, who film sports using drones. He agreed to take some photos of the hotel and nearby territory.
This way we got lot’s of overlapped photos and built simple Ortho. Applied our new model to the Ortho image. It works!
We added GTIFF support as “upload your own imagery” to Mapflow.ai 🎉
☝️ And this is one more important feature release on Mapflow.ai which is related to what is written all above.
Now you have three options to run the model through the Mapflow web app:
1) Select the default imagery source (Mapbox or ArcGIS)
2) Add Custom imagery URL (E.g. add the image from Openaerialmap)
3) Upload local GTIFF file
Enjoy the new model and share your feedback with us. We appreciate the comments on our blog or by email the most but we also encourage you to rate the model’s performance by scoring it in the map view. ⭐️⭐️⭐️⭐️⭐️