I wonder if this is a plausible problem. I get MAEs bellow 0.3 in validation, however, when predicting test images seems very difficult to get below 0.69, the same value that gives a submission with all 0s.
Just to make sure and not to waste time.
You have to put in $40,000 work to get your $4,000 ... this is my first time competing and I submitted a very simple network just to test and afterwords built a few more that I was sure would do better. The simple network still has the best score by far. It seems pretty difficult as you notice. Also those images!? They seem very low quality or maybe I am just not getting it, but I am unable myself to visually identify the fields. I'm using the supplied starter notebook to draw them but it does not make sense.
I think it's difficult in ML to generate the distance point relationship (especially with this data).
however, read this topic: https://zindi.africa/competitions/lacuna-correct-field-detection-challenge/discussions/5776