Request for Pretrained Weight File for Swin Transformer on Mask R-CNN

[aliwaqas333]

First of all, I would like to thank you for your outstanding work on integrating Swin Transformer with Mask R-CNN and providing the valuable resources and codebase to the community.

I have encountered some difficulties in achieving satisfactory training accuracy. My results are considerably lower than the ones in your paper. Here are a few key metrics from my training results:

Bounding Box Detection (bbox):

• AP @ IoU=0.50:0.95 (all areas, maxDets=100): 0.181
• AP @ IoU=0.50 (all areas, maxDets=100): 0.403
• AP @ IoU=0.75 (all areas, maxDets=100): 0.128
• AP @ IoU=0.50:0.95 (medium areas, maxDets=100): 0.191
• AP @ IoU=0.50:0.95 (large areas, maxDets=100): 0.214

Segmentation (segm):

• AP @ IoU=0.50:0.95 (all areas, maxDets=100): 0.104
• AP @ IoU=0.50 (all areas, maxDets=100): 0.343
• AP @ IoU=0.75 (all areas, maxDets=100): 0.030

My training configuration is as follows:

PYTHONFILE=scripts/mmdet/train.py
CONFIG_FILE=configs/mmdet/swin/mask_rcnn_swin-t-p4-w7_fpn_ms-crop-3x_coco.py

python ${PYTHONFILE} ${CONFIG_FILE} --work-dir ${WORKDIR}

The availability of the pretrained weight file would greatly enhance the reproducibility of your results.