COCO2017 rotated bounding boxes hyper parameters

[cameronchoi]

I've been trying to train for rotated bounding boxes on the COCO2017 dataset, but can't seem to recreate as good of results as shown in this blog post.

The best results I've gotten so far has been as shown below.

Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.06898
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.17053
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.03680
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.04298
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.08473
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.07430
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.13377
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.22606
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.23517
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.12647
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.23607
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.29366

For this I used the command:

odtk train rotated_model.pth --fine-tune /coco/pretrained/retinanet_rn50fpn.pth --images /coco/images/train2017/ --annotations /coco/annotations/instances_train2017_RBB.json --backbone ResNet50FPN --val-images /coco/images/val2017/ --val-annotations /coco/annotations/instances_val2017_RBB.json --batch 12 --rotated-bbox --lr 0.0005 --iters 250000 --val-iters 10000 --milestones 200000

I trained on 6 NVIDIA GeForce RTX 3090s.

I've also made sure to calculate theta for the annotations exactly as specified in the blog post.

1. What mAP score was able to be achieved for rotated bounding boxes on COCO2017 for the blog post?
2. What hyper parameters were used to achieve these results?