Any one has got reference for good object counting model that give fair count for occluded objects scene belong to different classes ?
Yolo will give poor results because of occlusion
Hi there,
From what I’ve read, two stage detection networks tend to work better for occluded objects. That also seem to be in line with what you’ve encountered from your experiments, as YOLO is a one stage network.
So perhaps try a two stage like Faster RCNN or RFCN and see if it improves?
Also, different so called occlusion techniques could be used as data augmentations to make the model more robust against occluded objects like shown here:
Perhaps also have a look at different SOTA models for object counting in specific from paperswithcode to get more ideas:
I don’t know what kind of objects you’re trying to count, but I’ve used this in a project with fairly good results for objects that are fairly similar to each other:
1 Like