Seems like segmentation would be a better approach to identity objects in a photo rather than various other features.