Some questions on annotation accuracy

[phflot]

Hello, I really like you work! - the huge amount of very difficult poses and quality makes your dataset unique - and we are considering to use your dataset as a benchmark in our research. Could you give us more details on how the frames have been annotated? Firstly, I believe that the 43 point landmarks seem to be the profile but the landmark id does not seem to be consistent, i.e. it is always the first 43 coordinates without a way to tell if it is left or right, or am I missing something? In the paper you write about manual annotation but also about a facial landmark detector and projection between rgb and thermal. Are those methods also used to generate ground truth landmarks? I encountered a few weird annotations, e.g.:

and I am wondering if I did something wrong when finding the correct dataset entry for a given image or how else to explain those images when removing them for evaluation.

Finally, would you be able to reconstruct annotations with subpixel precision (e.g. did rounding happen when you exported the annotations)? The quantization artefacts might make comparisons difficult, and having better annotations especially for the low res images would be incredibly useful:

Thank you for all infos!

[phflot]

Are there any updates on the issue?

[phflot]

push