Annotate images by some points. This policy makes the image rotatable in detection.
It took me 10 hours every day in 2 weeks to annotate all images. By the way, some annotation errors in the original VOC2007 dataset are fixed in our annotations. For a fair comparison, I only annotate the objects that are annotated in the original dataset.
We test the new annotations using SSD, which causes a 0.3 mAP drop in original images. However, if we rotate the images by 75 degrees, the mAP of SSD that train on original datasets drop to 20.0+, while the network that trained on our new data remains 40.0+ mAP.