Closed GloryyrolG closed 3 months ago
Thank you for your interest in our work. POCO, by design, estimates per-part uncertainty (see here).
We observe that training with per-part uncertainty helps in faster convergence. However, since most HPS methods are trained with real images with pseudo-GT annotations, the annotations are not very accurate (especially for hands and feet), hence the uncertainty estimation for these parts are also not accurate. Hence, we convert per-part uncertainty to scalar body uncertainty (see here) for practical use in downstream tasks.
Hi @saidwivedi et al.,
Thx for ur work. may i ask if the code/function of per-part uncertainty is currently available? thx & best,