Hello, i am working on stitching the pipeline, so i can have a broadcast frame input (of soccer of course), and homography matrix as output,
So i go by edge and pitch detection, works good,
and then feature creation using deep learning,
at this point, when i see the extracted features and the ground truth features (in test_features) i find it highly different,
Also if i continue, the extracted homography is somehow like 0.49 IOU with the gt_homography,
The only difference is the features, which should be extracted from the deep learning model trained in your repo,
So is there any parameters, any kind of tuning needed to be done to avoid this?
Hello, i am working on stitching the pipeline, so i can have a broadcast frame input (of soccer of course), and homography matrix as output,
So i go by edge and pitch detection, works good, and then feature creation using deep learning, at this point, when i see the extracted features and the ground truth features (in test_features) i find it highly different, Also if i continue, the extracted homography is somehow like 0.49 IOU with the gt_homography,
The only difference is the features, which should be extracted from the deep learning model trained in your repo,
So is there any parameters, any kind of tuning needed to be done to avoid this?
Thanks.