EgoCross Challenge at CVPR 2026: Cross-Domain Egocentric Video QA

The inaugural EgoCross Challenge took place during the Third EgoVis Workshop at CVPR 2026, unveiling a benchmark for egocentric video question answering across different domains. This challenge assesses the ability of multimodal large language models to extend their generalization beyond typical everyday situations by utilizing first-person videos from four specific areas: surgery, industrial assembly, extreme sports, and animal viewpoints. Each evaluation consists of an egocentric video, a corresponding question, and four possible answers for the model to choose from. The technical report outlines the challenge's tasks, benchmark resources, and two official Codabench tracks: the Source-Limited Track, which confines participants to the official baseline model and a limited support set, and the Open-Source Track, allowing more flexibility in model and training data selection while banning manual target-domain training data construction. The challenge seeks to enhance AI comprehension in specialized visual contexts, impacting sectors such as medical training, industrial automation, and wildlife research.