In the Keypoints class, the merge_with_descriptors method changes the dtype of the keypoints to that of the descriptors. If the descriptors were extracted with a low-precision dtype (e.g., float16 or bfloat16), this casting would cause keypoints extracted from megapixel images to shift by up to a full pixel.
disk = KF.DISK.from_pretrained("depth").to("cuda")
with torch.cuda.amp.autocast():
features = disk(megapixel_image, 1024)
print(features[0].keypoints.dtype, features[0].descriptors.dtype)
### Expected behavior
It might make sense to cast them to float32 or just leave their dtype unchanged
### Environment
```shell
N/A
edgarriba
commented
1 month ago
Describe the bug
This isn't exactly a "bug" but...
In the
Keypointsclass, themerge_with_descriptorsmethod changes the dtype of the keypoints to that of the descriptors. If the descriptors were extracted with a low-precision dtype (e.g., float16 or bfloat16), this casting would cause keypoints extracted from megapixel images to shift by up to a full pixel.https://github.com/kornia/kornia/blob/bdeac07e1edc26863a9c8d0826dc202974fd850a/kornia/feature/disk/structs.py#L82
Reproduction steps
Additional context
"Bug" found in commit bdeac0