Can't reproduce Faster RCNN Resnet50FPN mAP for coco2017

Hello I am using FasterRCNN model with the pretrained RESNET50 FPN backbone for 91 classes.
For my project I am using many models and one of them is YOLOv3 and because of this I chose to have the coco2017 dataset on YOLO format (txt file with the format < class:[0,79] >, < xc:[0,1] >, < yc:[0,1] >, < w:[0,1] >, < h:[0,1] > bounding box per line.)
When I try to compare Faster RCNN performance I am converting the labels from 80 to 91 and the coordinates to xmin xmax ymin ymax. After this I am training Faster RCNN with the pretrained 91 Renset50FPN backbone with a learning rate 0.001, SGD with weight decay=0.0005 and momentum=0 for 13 epochs. i am reducing the learning rate by a factor of 10 in epoch 8 and 11 but that does not make a difference.
This is my result:
I have seen that the baseline for FasterRCNN on coco2017 is 36% which is 3.4 percent bigger than my results.
I cannot understand why there is this difference. Do you have any ideas?
Thank you