strip_rcnn_s_dota.pth Checkpoint issue

[eperec03]

Dear authors,

Thank you for releasing Strip R-CNN. While evaluating the pre-trained
checkpoint (strip_rcnn_s_dota.pth) on DOTA v1.0, I encountered an
issue that may help you improve the reproducibility of the repository.

Checkpoint mismatch with current codebase

The downloaded checkpoint contains weights for conv_spatial1/2 in the
RoI head, while the current implementation (rotated_bbox_head.py)
uses strip_conv1/2. Additionally, the checkpoint is missing the
reg_theta_fcs.0/1 layers that now exist in the model.

PyTorch silently loads the checkpoint by skipping the missing keys and
initialising the new layers randomly, which causes a severe drop in
storage-tank AP (from ~72% to ~63% after reconstruction).

It would be helpful to either:

• update the official checkpoint to match the current main branch, or
• provide a legacy config that matches the released weights.