I was looking at this graph from Learning Data Augmentation Strategies for Object Detection and I noticed that the value for mAP is lower than all of mAP_S, mAP_M, and mAP_L for the third set of points. I would think mAP would be the weighted sum of those, but obviously not. Can someone tell me exactly how these are calculated? In the previous section (4.4), they mention using PASCAL VOC and therefore using an IoU threshold of 0.5, but in this section (4.5) they are using COCO, so I would assume they're using IoU thresholds of [.5:.95]. Is this correct?
I emailed the authors and they were kind enough to get back to me, so credit to them. The key thing to note here is that these are relative improvements, not actual values. So if you take a look at the COCO leaderboard, you'll see the top result is [0.407, 0.616, 0.720, 0.588] for [mAPs, mAPm, mAPl, mAP]. The second-best result is [0.399, 0.605, 0.706, 0.578]. In both of these cases, the value for mAP is within the others, but if you calculate the relative improvement of the best over the second-best, it's [2.0, 1.8, 2.0, 1.7], thus the mAP relative improvement is lower than all of the others.