They explained that the model over-counted attendance because the hexagonal rock formations appeared too similar to people when viewed from above.
“Object detection models rely on patterns like shape, texture, and contrast to recognise objects,” they added.
“Here the rocks and humans have similar contours, shadows, and colours, the model may misclassify the rocks as people.
“This happens because the training data may not include enough examples of the Giant’s Causeway or similar environments, leading the model to generalise incorrectly.”
They suggested that the AI approach would work in future if the model was trained with “considerably more data” and if more expensive high resolution drone filming was used.