Instance Segmentation using CPU fails on certain images when Swin Transformer backbone was used

[eric-kwok-nt]

The error that I have encountered when inferencing using the swin transformer on CPU was ERROR - upper bound and larger bound inconsistent with step sign

Such error disappears when inference was performed on GPU.

After some investigation it was found that when running on CPU, the bbox (batch size 1) that was provided to the function _do_paste_mask in the file mmdet/models/roi_heads/mask_heads/fcn_mask_head.py has negative coordinates, causing it to fail.