EgoDG-Bench: Egocentric Video Domain Generalization in the Era of VLMs
Abstract
Domain shifts in egocentric videos, due to factors such as changes in environment, scenario, or geography, remain an under-explored challenge. Although a few methods for video domain generalization (DG) exist, a comprehensive analysis of their effective- ness on these factors of domain shifts in egocentric videos is lacking, primarily owing to the absence of a unified evaluation benchmark. Thus, we propose Egocentric video Domain Generalization Benchmark (EgoDG-Bench), a dataset designed to investigate the effects of various types of distribution shifts in egocentric videos. We study the effectiveness of current state-of-the-art approaches for DG on this benchmark, and surprisingly, these methods offer only modest improvements over a lower-bound baseline ERM [1].
We then ask: how relevant is the problem of egocentric video DG in the era of vision-language models (VLMs)? To address this question, we perform a thorough analysis of VLM performance across 17 target domains of EgoDG-Bench, evaluating their zero-shot capability, the effects of linear probing and prompt tuning, and com- pare six VLMs with twelve DG baselines under three types of domain shifts. Our findings show that even the best-performing VLM, LaViLa, is susceptible to such shifts—exhibiting a substantial performance drop (e.g., from 42.19% to 22.97% in the USA → India scenario). We further validate the impact of domain shifts through a series of experiments. Our proposed EgoDG-Bench facilitates the most comprehen-
sive evaluation of DG in egocentric settings—an essential step toward enhancing the
robustness of VLMs in this field. We will release our code and the dataset to support
future research in this field.
We then ask: how relevant is the problem of egocentric video DG in the era of vision-language models (VLMs)? To address this question, we perform a thorough analysis of VLM performance across 17 target domains of EgoDG-Bench, evaluating their zero-shot capability, the effects of linear probing and prompt tuning, and com- pare six VLMs with twelve DG baselines under three types of domain shifts. Our findings show that even the best-performing VLM, LaViLa, is susceptible to such shifts—exhibiting a substantial performance drop (e.g., from 42.19% to 22.97% in the USA → India scenario). We further validate the impact of domain shifts through a series of experiments. Our proposed EgoDG-Bench facilitates the most comprehen-
sive evaluation of DG in egocentric settings—an essential step toward enhancing the
robustness of VLMs in this field. We will release our code and the dataset to support
future research in this field.
