in the metadata.csv, we are given metadata (gsd and city) for train and test images.
are we allowed to use them at prediction for the test images?
(Note: there is almost a direct relationship between gsd and the target vflow scale and angle values. I wondered is this an unintentional leak in the data. It is possible to give very accurate estimate vflow scale and angle using using only gsd and city and without using image at all )
Yes, you are allowed to use GSD and city at prediction time. Per the official rules, participants can use any data we have provided on the competition page - external data is not allowed.
We anticipate that predicting object heights will be the most challenging part of the competition, while predicting scale factor and angle will be relatively simpler even without using the relationship with GSD and city. Relying too heavily on a city-based exploit will also make a model less generalizable on new cities, which would be relevant for the write-up track.
If anything is still unclear, please just let us know - we’re happy to help. Good luck!
Thanks for the reply.
I will probably have 2 submissions, one for top ranking and one for generalisation (write-up prize).
It has always been a dilemma in data science competition: to get top ranking, one may have to exploit the data characteristics (ending up not necessarily the best generalization solution, but rather a solution that best overfits the test data).
“We anticipate that predicting object heights will be the most challenging part of the competition”. Actually, I note another problem. Now we are estimating AGL and not depth (from camera plane). AGL is fixed for the building regardless of the viewpoint/oblique angle change. Hence I think the CNN model is recognising the building and not the AGL.
For the write-up bonus prize, it is fine to describe a model that is not your top-scoring submission, so long as it corresponds to a qualifying score (i.e. in the top 15 on the private leaderboard that is published when submissions close).
You are correct that the aim of the challenge is to measure building height, not depth from the camera.