DapengFeng

@sltlls In my case, it works for the min_area_polygons test. ``` >>> import torch >>> from mmcv.ops import min_area_polygons >>> pred_pt = torch.tensor([[1.0350e-02, 1.0234e+03, 1.0242e-02, 1.0259e-02, 1.0291e-02, 1.0244e-02, 1.0240e+03, 1.0249e-02,...

The error occurred in judging the collinearity of the points. For example, as shown in the figure, the value of `sign = cross(point 6, point 7, point 1)` and `sign...

> ``` > >>> import torch > >>> from mmcv.ops import convex_giou > >>> device='cuda' > >>> pred_pt = torch.tensor([1024 for _ in range(18)], > ... dtype=torch.float32, > ... device=device).unsqueeze(0)...

Changing data type will relieve the number instability but not solve the problem. Pre-sorting the points in order guarantees that the newly added point is always outside the convex hull...

Thank you for review. I will check them later.

Does it mean to replace Eigen with libtorch?

> Are you working on this? Yeah, I will try it.