I don't know how many others encountered the same issue but it took me a while to get correct predictions from the model. Anyway, I'm living this here hoping it can help someone.
So, I wanted to create some code to add my support set (not from COCO or PASCAL) and test the performances on some query images. For some reason, the final bounding boxes were misplaced entirely with respect to the instance. Later I found out that the correct points were lying on the line, parallel to the secondary diagonal, passing through them, mirrored with respect to the main diagonal. Now, I am fairly sure I caused the problem myself by not processing the input correctly but I patched it geometrically and now it works.
I don't know how many others encountered the same issue but it took me a while to get correct predictions from the model. Anyway, I'm living this here hoping it can help someone.
So, I wanted to create some code to add my support set (not from COCO or PASCAL) and test the performances on some query images. For some reason, the final bounding boxes were misplaced entirely with respect to the instance. Later I found out that the correct points were lying on the line, parallel to the secondary diagonal, passing through them, mirrored with respect to the main diagonal. Now, I am fairly sure I caused the problem myself by not processing the input correctly but I patched it geometrically and now it works.
Here's the full code:
And this is the test image I was using: "https://media.ktoo.org/2013/10/Brown-Bears.jpg"