Is it possible to derive the level of occlusion(reasonable/Heavy or visibility percentage) of the detected pedestrian in Crowd? Could you please mention how? about pedestriandetection-hgpd HOT 3 OPEN

CityPersons and CrowdHuman are two popular datasets with focus on the crowd occlusion issuse. Both of them provide visible box and full-body box annotations, thus you can derive the visibility ratio by calculating the overlaps between visible and full-body boxes: visibility ratio = Area(visible) / Area(full-body). Or you mean taking visibility ratio as the regression target during the training?

from pedestriandetection-hgpd.

Thanks for replying. I would like to mention:

1. Suppose I do not have the visible bounding box, then can I derive occlusion level with the full-body bounding box only?
   For example, by calculating MaxIoU ( the max IoU with other same category objects in a single image. source: https://github.com/liruilong940607/OCHumanApi)

2. If I have a visibility score and related full body bounding box, then Can I train this model to regress the visibility score when an object is inputted? I think yes, I meant taking the visibility ratio as the regression target during the training. As my goal is to input a pedestrian's full-body bounding box and estimate its occlusion level.

from pedestriandetection-hgpd.

Thank you for your sharing!

1. I don't think we can obtain the accurate occlusion ratio only with full-body boxes. I take the MaxIoU as an example. There are two overlapped persons, of which one is the occluder, and the other is the occludee. The occluder may be fully visible, but it has a high MaxIoU with the occludee. Therefore, the MaxIoU fails to indicate the occlusion ratio for occluders. Relative occlusion order is essential to accurate occlusion levels. In other words, we need to figure out which is the occluder (or occludee).
2. Predicting the accurate visibility ratio is a difficult task. Estimating the occlusion level has been researched in previous works (Beta R-CNN, Bi-Box, MGAN), however, they usually predict the visible box or visible mask to indicate the occlusion level, instead of directly predicting a scalar (visibility ratio).

from pedestriandetection-hgpd.