The traditional approach

For the first two years of the ImageNet competition, the entries all looked roughly the same.

The dominant approach to computer vision at the time was hand-crafted. Researchers would sit down and decide, carefully and manually, what the system should look for: edges at specific angles, patterns in texture, the distribution of colors in a region of the image. They would these rules into the system by hand. The system would then check incoming images against them.

It was the old expert system logic applied to vision. Humans deciding what matters. Machines following instructions.

The approach worked, slowly and incrementally. Each year brought small improvements. Enormous effort for marginal gains.

The 2011 winner classified images with roughly 26% error: given five guesses, the system was wrong more than a quarter of the time. That was state of the art.

The traditional approach

For the first two years of the ImageNet competition, the entries all looked roughly the same.

It was the old expert system logic applied to vision. Humans deciding what matters. Machines following instructions.

The approach worked, slowly and incrementally. Each year brought small improvements. Enormous effort for marginal gains.

The 2011 winner classified images with roughly 26% error: given five guesses, the system was wrong more than a quarter of the time. That was state of the art.