Source: Deep Learning on Medium
Pose2Seg: Detection Free Human Instance Segmentation
The pose can be used → to help instance segmentation → this is much better than regions since it can handle occlusion. (they present new dataset as well) → this is very smart research. (occlusion is now on the solve list).
There are real-life applications when it comes to human segmentation → and now we can take advantage of good prior → which is a human pose. (use pose to perform segmentation).
Most of the methods → region mask → and do non-maximum suppression. (but when there is an overlap → thing are going to get removed).
Human is a special category → and we need to take into account. (and there are a lot of application → around this technology).
So → this method is using human pose rather than region.
This is good → and the model has a good translation layer → to undo the image transformation. (very sexy research).
How the author’s method is done. (good amount of contributions).
This method can be really powerful! (most of the methods create bounding boxes → and then start to reduce the overlaps) → however, for humans, there can be occlusion. (how can we solve that better).
A new dataset has been released.
Multiple pipelines → we are going to have to have the skeleton. (but very powerful framework). (also, they trained on COCO and tested on the new dataset).
Reset is used to perform segmentation. (the alignment is done via → human pose rather than bouding boxes).
There is a k-means clustering happening within the layer → super complicated.
Much clear segmentation is → very powerful. (they double the accuracy) → which is very impressive.
Very good segmentation results in occlusion. (some of the translation we see in computer vision → really hard to undo or solve). The only downside is that we need some human pose to use this model → would be good if that can be removed.