Adaloss: Adaptive Loss Function for Landmark Localization

Source: Deep Learning on Medium

Adaloss: Adaptive Loss Function for Landmark Localization

Another loss function for keypoint detection → but now we are using a heatmap → we are going to use a 2D probability map. (but this loss is special since it adapts to the ground truth data → and performs gradient control)

So certain loss values are more focused → such as small or medium losses. (for any computer vision-related task → there are some kinds of key points we need to detect → , however, actually predicting these things are hard to do).

Impressive outcomes → very precise → very accurate. (most existing deep learning method → outputs coordinates or heatmaps → and they are using)

The 2D Gaussian is spread out at first → but as training goes → it becomes more precise → this is good since the network is working on a harder problem. (at first, → we are going to use a lot of variances → so the training can happen much easier → however → as the training proceed → we are going to get more precise. )

This paper is special since → they create an application based → adaptive loss function → which is super interesting. (it needs fewer iterations for the model to converge → this is good).

They tried their method on → different applications → hopefully they will be able to show an impressive results for all of them. (are heatmaps better? → it seems so → after adding skipping connections as well as FCN) → some works are done related to videos and they are data augmentation skims) → some have even used Gans to create a new dataset → this is a hard task to do.

And they are using the heatmap regression method → and their loss function is specially created → where the optimizers are thought of before creating the loss functions. (this is such an interesting approach of creating an adaptive loss functions → others do it in a way there are some hyper-parameter we need to set → and depending on those values the gradients change → not tweaking the optimizer).

And as seen above → the loss value becomes much smaller → this is because the Adaloss is able to adapt to current error loss value → making things much more harder → for better precision.

They used different dataset → medical images as well as cat images.

Adaloss is really good → better convergence properties. (impressive) → the distance becomes much more precise.

The loss value → becomes → much smaller → precise.

Here we can see that the standard deviation of the Gaussian distribution becomes smaller and at convergence, it does not decrease.

Wow, it is very impressive → we are able to set a sigma value for certain body parts → the ability to focus on certain body types seems like a great idea.

Eyebrows and Jawlines are hard to predict since they might be covered or due to the angel there it is not shown.

We are able to observe that the author’s method gives many stable results → very impressive! (and using this method for different applications such as Medical Surgery is another great application).

And they were able to get good results → this is another important find → since these technology is able to give an open door for e-commerce.

They are able to advance the state of the results → thanks to creating a new loss function where the problem becomes much more hard as the training continues on. (gradient control)