xxxmy

what's the dataset you used? And have you drawn results on images? The eval code is mostly same as pyfastrcnn.

Hi, I found that you created your branch before I fixed a bug of preprocessing.

the official pre-trained weights of resnet contains fc layer that we don't need. Therefore, you can modify line 197 as below: `model.load_state_dict(model_zoo.load_url(model_urls['resnet50']),strict=False)`

@Single430 看训练代码作者用的是voc07metric，voc_eval.py应该用的是voc10的标准，两个AP计算方式有不同，但是R这么低应该是有问题的