mAP for ssd and ssdlite caffe models without any finetuning?

[fcakir]

Does the repo owner or any others have accuracy values for the ssd and ssdlite caffe models with the tensorflow weights?
The mAP I get with ssd mobilenetv2 is 0.023, which is significantly lower than the official result. This is even when using relu6.
Has anyone successfully reproduced the official mAP values?