Why can we not use the weights learned in the VOC object detection? It seems like this would allow for much better accuracy than using weights that start from ImageNet. Would love to hear if this i...