Sports Field Homography is designed to predict the homography of sports fields such as a basketball court or soccer/football pitch. This repository implements an end-to-end model consisting of the UNET network, the output of which is connected to the Spatial Transformer Network (STN).
wonderful work!
I have read some papers about homography matrix estimation, eg: Content-Aware Unsupervised Deep Homography Estimation or Homography Net, they estimate offset vector for 4 vertices. While you directly regress a homography matrix consisting of 9 floating point values.
So I wonder have you compared the effects of the two?
Thank you!
Best wished!
When calling ./scrips/run_inference.sh --game=...
some files in subfolder assets/pretrained are missing - this is the error: FileNotFoundError: [Errno 2] No such file or directory: '/home/.../test/sports-field-homography/assets/pretrained/ncaav8-640x360-aug_unet-resnet34-deconv-img+mask_ce-l1-rrmse-focal_pre/last.pth'