Closed YoussefFathi closed 3 years ago
The uploaded model is only for axis-aligned bounding boxes. We don't have any pretrained models for rotated objects because labels need to be created for this purpose. You can check out this blog: https://developer.nvidia.com/blog/detecting-rotated-objects-using-the-odtk/
Hi @yashnv
Is there any out of box utility that creates a rotated bbox annotation from mask images?
The task - i already have images and annotations for an instance segmentation task, i need to evaluate retina-net for the same to compare against rotated maskrcnn and yolact.
@ghost before I open a new issue, that means no model was trained from scratch with rotated bb? In the blog, they used rotated bounding boxes only for transfer learning the model?
Is there a pretrained model available that was trained on rotated bounding box detection and can be used directly for inference of rotated bbox? I've tried using the uploaded pretrained models (i.e ResNet18FPN) for inference with (--rotated-bbox) however I get this error:
RuntimeError: Error(s) in loading state_dict for Model: size mismatch for cls_head.8.weight: copying a param with shape torch.Size([720, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([2160, 256, 3, 3]). size mismatch for cls_head.8.bias: copying a param with shape torch.Size([720]) from checkpoint, the shape in current model is torch.Size([2160]). size mismatch for box_head.8.weight: copying a param with shape torch.Size([36, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([162, 256, 3, 3]). size mismatch for box_head.8.bias: copying a param with shape torch.Size([36]) from checkpoint, the shape in current model is torch.Size([162]).