Case Study: Yandex achieves faster, higher-quality, 10× cheaper image labeling to train self-driving cars with Toloka

Toloka Helps Train a Self-driving Car to Detect Surrounding Objects

Yandex, which has developed an autonomous vehicle fleet of 200 self-driving cars that have logged over 4 million autonomous miles, needed tens of thousands of accurately labeled images to train its object-detection neural networks for Russian city streets. Open datasets didn’t match local road conditions and buying labeled images was costly (~$4 each), so Yandex used Toloka to cost-effectively scale image labeling for its autonomous driving pipeline.

Toloka delivered an API and embeddable interface that let Yandex integrate a custom visual editor (layers, transparency, selection, zoom, classification), automatically split tasks, and combine human “Tolokers” with neural-network labeling plus cross-check verification. This approach sped up and improved label quality, integrated Toloka into Yandex’s ML workflow, and reduced labeling costs by about tenfold compared with purchased labels.

Yandex