The resault of eval.py is strange

I download the trained model you provide and use txt2xml.py to change the label file format(the original txt2xml.py doesn't work on my computer, so I rewrite some code about reading the label file). When I use the eval.py to the DOTA train dataset, I just have a very strange result, the mAP is just about 0.2. Recall of helicopter and small-vehicle is below 0.1. When I use demo.py to visualize the result of some train, the result is not that bad.