In general, IoU calculation for rotated boxes [1] is not differentiable as it requires triangulation to calculate the area of intesection polygons, like the box_iou_rotated op.
[2] proposes an IoU Loss for rotated 2D/3D objects, however, both implementation details and codes are not provided. [3] shows to calculate IoU between polygons in a pixel-wise manner.
Recently, we find two implementations for differantiable IoU calculation, [4] and [5].
Here we use the picture in [4] to illustrate this method.
- Find intesection points of two polygons.
- Sorting these points to get the indices.
- Using
torch.gatherto fetch the real values of points. - Calculating intesection areas by Shoelace_formula [6].
As the torch.gather is differentiable, the whole function is differentiable.
We modify the code in [4] to our codebase and evaluate its performance on DOTA and HRSC2016. Specifically, we add an op called box_iou_rotated_diff lies in here.
The results show that optimizing bbox regression with IoU Loss can further boost the performance. Here we list the results in the table below.
| Model | Data | Backbone | reg. loss | Rotate | Lr schd | box AP |
|---|---|---|---|---|---|---|
| RetinaNet | HRSC2016 | R-50-FPN | smooth l1 | ✓ | 6x | 81.63 |
| RetinaNet | HRSC2016 | R-50-FPN | IoU | ✓ | 6x | 82.74 |
| CS2A-Net-2s | DOTA | R-50-FPN | smooth l1 | - | 1x | 73.83 |
| CS2A-Net-2s | DOTA | R-50-FPN | IoU | - | 1x | 74.58 |
References
[1] Arbitrary-oriented scene text detection via rotation proposals
[2] IoU Loss for 2D/3D Object Detection
[3] PIoU Loss: Towards Accurate Oriented Object Detection in Complex Environments
[4] https://github.com/lilanxiao/Rotated_IoU