Hi, I want to use NVIDIA’s PlaneRCNN. The code is old, and I guess it belongs to the pre torchvision era. Some parts of code need nms
and roialign
functionality but to use them, some compilation should be done. I tried that, and there is a lot of error in windows, and it turns out this is a known issue.
As a workaround, I want to use torchvision nms
and roialign
functionality. I somehow implemented expected nms
for the code as follow:
import torchvision
import torch
def nms(dets, thresh):
iou_threshold = thresh
x1 = dets[:, 1][..., None]
y1 = dets[:, 0][..., None]
x2 = dets[:, 3][..., None]
y2 = dets[:, 2][..., None]
boxes = torch.concat([x1, y1, x2, y2], axis=1)
scores = dets[:, 4]
return torchvision.ops.nms(boxes, scores, iou_threshold)
I don’t know if it is working, but I think it is ok. For the roialign
part, there is an import line as follow:
from roialign.roi_align.crop_and_resize import CropAndResizeFunction
I want to know whether there is any viable solution in torchvision to implement this. Any further help and suggestion about the approach are highly appreciated.