Pedestrian Detection - Still Banging Our Heads

I've been not following autonomous driving and computer vision for some time already. So I am astonished that we are still banging our heads against the wall.

I might not be a fan of autonomous driving disrupting industries (though I'd like to have an option to give away the wheel on a highway ;) ). At the same time even beyond autonomous driving, pedestrians detection technology holds value for intelligent transportation. It's disappointing to discover that the industry is stubborn to use CNNs for this purpose.

In 2024 Peking University and University College London conducted research that shows that while overall detection accuracy of pedestrian detection by the newest models is up, the ability to detect children is getting worse. This isn't just a dataset problem. Humans can spot a person in any context because our brains don't rely on brute-force data training like convolutional neural networks.

We're pouring money into refining systems that will never match human-level safety because they learn in a fundamentally different way. In a way that is by definition prone to amplifying data biases both in over-representation and under-representation. What if we invested in understanding and mimicking how the human brain actually learns and adapts? It seems that the industry is not capable of changing course, though.

And this is only about detection. We must not forget that still the possibility of adversarial attacks on computer vision is a huge problem. Even Geoffrey Hinton admitted he was wrong saying deep learning will be able to do everything.