God’s Favourite.
God’s Favourite.
ML coordinate prediction in visual data. A model that learns where to look.
The premise is simple: given an image, predict a coordinate. The execution is not. An EfficientNet-b0 backbone handles feature extraction, with a regression head outputting (x, y) predictions. The training signal comes from an unexpected source — human gaze data, aggregated across thousands of responses, used to teach the model what humans actually look at versus what a naive detector would select.
The gaze overlay is the differentiator. Raw image features get you in the ballpark. The human attention signal — where people actually focus when they look at the same image — narrows the prediction to a point the model could not reach on its own. The gap between “what the image contains” and “where humans converge” is the entire value of the system.
Won on the first entry. Six competitions completed with active iteration on the model architecture, training pipeline, and feature engineering. The feedback loop is tight — compete, score, adjust, re-enter.
The premise is plain: given an image, predict one coordinate inside it.
A pretrained backbone extracts the features, and raw features alone get the prediction into the ballpark.
A regression head turns those features into a single (x, y) point.
I train against human gaze, aggregated across many responses, so the model learns what people actually look at rather than what a naive detector would pick.
The gap between what the image contains and where people converge is the entire value of the system.
The prediction goes in as an entry.
Every entry comes back scored, which keeps the feedback loop tight.
Between entries I iterate on the architecture, the training pipeline and the features, then go back in.