Hybrid incremental learning of new data and new classes for hand-held object recognition.

Highlights • We have polished the result analysis in the experimental section for clarification. • We have added missing references to make the references more comprehensive. • We have carefully gone through the paper and did a thorough revision, including grammatical errors, typos and the formats. Abstract Intelligence technology is an important research area. As a very special yet important case of object recognition, hand-held object recognition plays an important role in intelligence technology for its many applications such as visual question-answering and reasoning. In real-world scenarios, the datasets are open-ended and dynamic: new object samples and new object classes increase continuously. This requires the intelligence technology to enable hybrid incremental learning, which supports both data-incremental and class-incremental learning to efficiently learn the new information. However, existing work mainly focuses on one side of incremental learning, either data-incremental or class-incremental learning while do not handle two sides of incremental learning in a unified framework. To solve the problem, we present a Hybrid Incremental Learning (HIL) method based on Support Vector Machine (SVM), which can incrementally improve its recognition ability by learning new object samples and new object concepts during the interaction with humans. In order to integrate data-incremental and class-incremental learning into one unified framework, HIL adds the new classification-planes and adjusts existing classification-planes under the setting of SVM. As a result, our system can simultaneously improve the recognition quality of known concepts by minimizing the prediction error and transfer the previous model to recognize unknown objects. We apply the proposed method into hand-held object recognition and the experimental results demonstrated its advantage of HIL. In addition, we conducted extensive experiments on the subset of ImageNet and the experimental results further validated the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]