Yiwen Chen

Current vision-language models can identify context in driving videos but fail to accurately assign fault in accidents, revealing a critical gap in autonomous driving AI capabilities.

Transforming 3D relighting into a token-space operation allows for intuitive, progressive lighting edits without the overhead of traditional methods.

Achieve globally consistent 4D radar maps by jointly optimizing geometric, inertial, and ego-velocity constraints, surpassing the accuracy of radar-inertial odometry and pose-graph methods.

Semantic IDs, which drastically reduce ID cardinality while inducing semantic clustering, have demonstrably improved ranking and retrieval in Snapchat's production recommender systems.