Vision Myths

Computer vision attracts a lot of confident claims. Some of them have cost companies millions. Before you finish this course, let's take on the most common myths, and replace them with what really happens on real projects.

• Confident claims
• Some cost millions
• Replace myths with reality

Myth: more data always helps

People say that more data fixes everything. It helps, but only if it's the right data. A million photos taken in good light teach your model nothing about bad light. Ten thousand varied images often beat a million similar ones. So before collecting more, look at where the model fails, and collect exactly that.

Myth: high accuracy means it works

A model scoring ninety nine percent can be useless. If defects are rare, always saying fine gets you that score. If the test set shares images with training, the number means nothing. And if the test data came from one week in summer, you've learned nothing about winter. Accuracy is a question, not an answer. On what data, and by which measure?

Myth: deep learning replaced classic vision

Deep learning dominates recognition. But geometry is still classic. Calibration, homography, and stereo are all classic methods, and they give exact answers. Classic tools run in microseconds, on cheap hardware, and need no training data. Most real systems mix the two. A network finds the object, and geometry measures it.

Myth: AI sees like we do

It's tempting to think that a model sees what we see. It doesn't. A famous study showed that image networks rely heavily on texture, so a cat with elephant skin gets called an elephant. Tiny changes, invisible to us, can flip a prediction. Models find whatever pattern's easiest, whether or not it's the one that you intended. So never assume that a model understands.

Summary