Dataset and Baselines for IID and OOD Image Classification Considering Data Quality and Evolving Environments

• DOI: 10.9781/ijimai.2023.01.007
• Publication Date: 2023-02-01T11:11:52Z
• Authors: Zhuo Zhang; Yang Li; Yicheng Gong; Yue Yang; Shukun Ma; Xiaolan Guo; Sezai Ercisli
• Keywords: evolving environments; Technology; T; active learning; efficient dataset; data quality; IJIMAI; generalization

At present, artificial intelligence is in a period of rapid development, and deep learning has begun to be applied various fields.Data, as key part the learning, its efficiency stability, will directly affect performance model, so it valued by people.In order make dataset efficient, many active methods have been proposed, containing independent identically distribution (IID) samples reduced with excellent performance; more stable, should solved that model encounters out-of-distribution (OOD) improve generalization performance.However, current method design adding OOD lack guidance, people do not know what selected which added better performance.In this paper, we propose variety elements called Complete Sample Elements(CSE), labels such rotation angle distance addition common classification labels.These can help analyze characteristics each element an efficient dataset, thereby inspiring new methods; also construct corresponding test set, only detect but helps explore metrics between existing guide samples, efficiently.In datasets terms element, confirm tends contain different appearance.At same time, experiments proved positive influence on dataset.