For the images captured with large camera rotation angles, the matching performance is much worse than that of SuperPoint.

I used the inference code,
```bash
CHECKPOINT=assets/models/silk/coco-rgb-aug.ckpt
TOPK=500
SIZE=640

./bin/silk-features -o -m pvgg-micro -c $CHECKPOINT -k $TOPK -d $OUT/features -s $SIZE $IMAGES/*.$IMAGES_EXTENSION
# generate image pairs to match
ls $OUT/features/*.pt | sort -V | xargs ./bin/generate-matches exhaustive -s > $OUT/matches.txt
# match keypoints from generated image pairs
./bin/silk-matching -o -m double-softmax -t 0.9 $OUT/matches.txt $OUT/matches/
# visualize matches
./bin/silk-viz image -o $OUT/viz $OUT/matches/*.pt
```

For images captured after a large rotation angle of the camera, the matching performance is much worse compared to SuperPoint.
For euroc dataset, matching result of silk model as follow:

matching result of superpoint model as follow:

Reproduce the reported case with the provided silk-features, generate-matches, silk-matching, and silk-viz commands on images with large camera rotations, using the stated checkpoint and settings. Compare the resulting matching quality with the referenced SuperPoint result; done requires identifying the cause of the gap and demonstrating improved large-rotation matching.

Written by the indexing model from the issue text.