How Alibaba’s AI Beats the KITTI Benchmark and Revolutionizes Visual Shopping

At CVPR 2017 in Honolulu, Alibaba’s “foot‑shopping” technology attracted attention.

A Google engineer visited Alibaba’s booth and used Taobao’s “拍立淘” feature to scan a person’s shoes; the system accurately identified the same shoe model.

“拍立淘”, launched in 2014, lets users tap the camera icon in the Taobao app, point at a product, and instantly find matching items.

During the China AI Conference (CCAI 2017) in Hangzhou, Alibaba showcased visual‑shopping demos such as “拍立淘” and a virtual try‑on mirror, drawing many visitors.

In May, Alibaba iDST visual computing researcher Hua Xiansheng’s team set a new world record on the KITTI vehicle‑detection benchmark, raising accuracy to 90.46 %.

The team’s solution combines region‑fusion decision making with a context‑aware multi‑task deep neural network to handle multi‑view, pose variation, and occlusion in complex scenes.

Alibaba’s advertising platform also achieved the best result in the ICDAR Robust Reading competition with its OCR technology, far surpassing the runner‑up.

Today the OCR system powers real‑time image review across the entire Alibaba advertising business, processing tens of millions of images daily with over 95 % accuracy and reducing risk‑detection time from days to seconds.

These advances illustrate that computer‑vision research continues to face challenges but also offers abundant opportunities, especially when combined with deep learning and big‑data techniques.