I stuck with this issue last week after reading your report on how to incorporate the robust loss.
For that the current methods should have MLE loss but aparently the current loss are not MLE based .
Check also its github and see the code arrangement and the dataset.
Most of the human pose estimation use heat map based methods but it has high complexity and requires high storage space.
In contrast, regression methods are used but they have inferior performance.
In this paper human pose regression is done based on maximum likelihood estimation for modelling the output distribution.
In this paper they claim that regression performance can be improved if we construct the likelihood function with the true underlying distribution instead of the inappropriate hypothesis
They propose a novel and effective regression paradigm, named Residual Log-likelihood Estimation
(RLE), that includes normalizing flows to estimate the underlying distribution
To be brief, Instead of fitting the original distribution they let the flow model to learn the change of distribution. They do not set the form of distribution in advance.
With Reparameterization design, flow model learns distribution which reflects the output deviation from the ground truth
Initially all the underlying distributions are assumed to be laplace distribution and the standard l1 and l2 loss are replaced by Residual Log-likelihood Estimation.
The optimal distribution is based on simple distribution, residual log term and the constant term which makes sure that the residual log term is a distribution.
-By using the above method they try to fit the underlying residual likelihood instead of learning the entire distribution. Finally, combining the reparameterization design and residual log-likelihood estimation they formulate the total loss function.
Hence in this paper they leverage the normalizing flow model to learn the residual log-likelihood w.r.t. to the
tractable initial density function and is found to work better than the heat map based methods.
Browse through this paper and the video.
https://jeffli.site/res-loglikelihood-regression/
I stuck with this issue last week after reading your report on how to incorporate the robust loss. For that the current methods should have MLE loss but aparently the current loss are not MLE based .
Check also its github and see the code arrangement and the dataset.