Concave shapes in output tree crown geometries

[CiSong10]

Dear developer,

While using Detectree2, I’ve noticed that most tree crown segments look nice and round, but some segments have weird, specifically, "concave" shapes. I noticed them because I think tree crowns are usually expected to be convex or rounded, while these outputs are concave. I’ve attached a few examples below to illustrate what I mean. There are only a few in one run of prediction, but there are always a few in every run, which I find interesting.

Although I may mitigate these cases by providing more training data, setting more strict confidence score and cleaning thresholds, etc. I am interested in if these imply something deeper about the model's behavior, such as training data contamination, or a specific model behavior caused this, etc.

Thank you for your time and for maintaining this excellent tool!

[PatBall1]

Hi @CiSong10 , yes this is something we've experienced from the very start. Some of the straight lines may be due to tiling artefacts (either in training or prediction) - we usually just hope these are filtered out. There are often more unusual jagged geometries which are difficult to interpret and I think are some kind of artefact of the masking networks. Do you have any other hypothesises that could be tested?

[CiSong10]

Thank you for your response! I do not have any hypothesis for now, but I will keep an eye on it