Yes, but historically this research is barely seven years old (Alexnet was published in 2012). There was lots of research into object detection before deep learning became usable, mostly based around feature engineering (keypoints, deformable parts, hog classifiers, etc).
Fundamentally the shift is that we've gone from handmade features developed over years of research to simply letting the model itself learn what's important. It would be nice to set the scene to realise just how transformative deep learning has been.
Yes. They provide this in the R-CNN paper, "Compared to the multi-feature, non-linear kernel SVM approach, we achieve a large improvement in mAP, from 35.1% to 53.7% mAP, while also
being much faster".
Also in the metrics provided in the paper, the best method was SegDPM had around 40.4% mAP. Beat it by a pretty large margin at its time.
