Home

Awesome

Keypoints Tracking via Transformer Networks

Model for sparse keypoints tracking across images using transformer networks

Paper : https://arxiv.org/abs/2203.12848

Our approach is hierarchical since a coarse keypoint tracking is accurately refined by a second transformer network. The model can be be used for both: image matching, and keypoint tracking

Architecture

<img src="./media/arc2.png" width="640" height="200">

Demo

Image matching

Match two image using SuperPoint descriptors (Basically tracking keypoints extracted by SuperPoint )

For example :

python demo_2_im.py --image1_path ="./media/im1.jpg" --image2_path="./media/im2.jpg"

alt text

Point Tracking

Tracking the points specified in demo_point_tracking.py.

alt text

Comparing with SuperGlue

Since there are no analogous works, we provide a comparison with the state-of-the-art method for keypoint feature matching SuperGlue. We tested our model on 2 datasets: COCO2014, and HPatches under different levels of light intensity, and projective transforms. For the evaluation metrics we selected 1) matching accuracy: the match is correct if distance between predicted and real position of the point is less than 6 pixels, and 2) number of correctly matched keypoints.

Results for acc.SuperGlue accuracyOur model accuracy, only coarse moduleOur model accuracy with fine module
COCO test easy homographies94.8%93.5%95.3%
COCO test hard homographies91.6%90.2%91.7%
COCO test hard homographies + illumination changes89.0%87.2%88.8%
Hpatches hard homographies91.4%90.5%91.5%
Results for number of correctly matched pointSuperGlue, number of correctly matched points out of 512Our model, number of correctly matched points out of 512
COCO test easy homographies249358
COCO test hard homographies240346
COCO test hard homographies + illumination changes195300
Hpatches hard homographies222340