- Field localization from a single broadcast image
- Early work using Deep Learning to directly estimate a Homography Matrix
Homography Optimization w/ Template
Given an homography estimate a registration error network is used to estimate the loss of the estimate . Since the registration error network is differential, it can be used to optimize .
Initial Registration Network
A ResNet-18 Architecture with the classification head swapped out for a fc-layer with an output of 8 (elements of the homography matrix).
Registration Error Network
ResNet-18 + Spectral Normalization +6-channel input (concat of input image& warped target template).
Trained from scratch.
- Random crop (light, 90% ~ 100% of original remains)
- Random horizontal flips
- Shadow simulation (random 50% opacity black patches)
- Random Translation (~1/2 full width), rotation (0~45 deg) and scaling(0.5~2x)
- Gaussian blur with kernel size 9
Interesting idea but is probably very compute intensive since a single iteration requires DNN inference. However, the optimization idea can be modularized and used as a post processing step along with ideas.