Pascal VOC2007 performance is not as good

[TheShadow29]

Hi. First of all thanks for your clean and clear repository.

I am trying to evaluate the models on VOC2007, however, I am unable to reach the required performance. SSD with resnet50 backbone on pascal voc achieves 79.7% map (reported here:https://github.com/ShuangXieIrene/ssds.pytorch#performance). However, in my experiments it is always saturating at 55mAP. I also tried with resnet18 and resnet34. The first saturates around 50, the second goes upto 55 but saturates thereafter.

The only possible difference is the input image size. I believe you are resizing images with image size of 608. I highly doubt that is indeed causing problem.

Would like to know your thoughts on this.

Also could you share your parameters to train on coco2017? In particular, how many epochs did you run it for.