Fast and memory efficient transformers based feature matching model.
This is transformers based Deep Learning Nework that is trained to match 2 sets of sparse image features. We propose methods to significantly reduce time spent for GNN, and completely remove Optimal Matching Layer while preserving similar matching accuracy with SuperGlue
For image matching run:
- python demo_2_im.py -v --image1_path= path_to_im1 --image2_path= path_to_im2
- python demo_2_im.py -v --image1_path='./media/im1.jpg' --image2_path='./media/im2.jpg' (exampe)
The resulting image will be saved in ./result
Result:
For video-image matching run:
- python demo_video.py --video_path=path --image_path=path
- python demo_video.py --video_path='./media/vid_car.MOV' --image_path='./media/ref_car.jpg' (exampe)
If you want to incude visualization of matching :
- python demo_video.py -v --video_path=path --image_path=path
The resulting video will be saved in ./result
| Number of keypoints per image | SuperGlue inference time (ms) | Our model inference time (ms) |
|---|---|---|
| 128 | 41.1 | 10.1 |
| 256 | 41.0 | 10.2 |
| 512 | 43.1 | 10.2 |
| 1024 | 55.7 | 10.3 |
| 2048 | 169.6 | 35.4 |
| 512 CPU | 180.5 | 35.4 |
