Core Concepts
DOZE introduces a dataset for Zero-Shot Object Navigation in dynamic environments, addressing the limitations of existing datasets and showcasing the potential for improvement in ZSON methods.
Stats
"DOZE comprises static and dynamic humanoid obstacles, open-vocabulary objects, objects with distinct spatial and appearance attributes, and hint objects."
"DOZE stands out by incorporating dynamic obstacles, open-vocabulary objects, and objects with distinct attributes."
"DOZE features scenes with an average of 8.62 diverse open-vocabulary objects, 52.45 objects with unique locations, and 48.24 with unique appearances."
Quotes
"Our dataset could be found at https://DOZE-Dataset.github.io/"
"Existing datasets for developing ZSON algorithms lack consideration of dynamic obstacles, object attribute diversity, and scene texts."