The ground truth metadata json's denote the bounding box dimensions as 'box': [697, 1451, 5720, 2649], but not this repository nor the fMoW paper describes what these dimensions mean. You have to search the baseline code and find in this function that these represent [x, y, w, h] where [x, y] are the upper (?) left corner's indices. Other bounding box annotation schemes exist, e.g. see here, so I think this would be something nice to have clearly documented somewhere.
Hi @rubencart,
You are correct, the order is "left, top, width, height".
I agree on the lack of documentation. Unfortunately this repository does not seem to be maintained anymore.
The ground truth metadata json's denote the bounding box dimensions as
'box': [697, 1451, 5720, 2649]
, but not this repository nor the fMoW paper describes what these dimensions mean. You have to search the baseline code and find in this function that these represent[x, y, w, h]
where[x, y]
are the upper (?) left corner's indices. Other bounding box annotation schemes exist, e.g. see here, so I think this would be something nice to have clearly documented somewhere.