Open World Object Localization Based on Negative Sample Learning

• Published in: Journal of Multimedia Information System Vol. 12; no. 1; pp. 13 - 26
• Main Authors: Yuan, Ling, Wang, Sanquan, Zeng, Lingke, Chu, Jun
• Format: Journal Article
• Language: English
• Published: 한국멀티미디어학회 31.03.2025
• Subjects: 컴퓨터학
• ISSN: 2383-7632; 2383-7632
• DOI: 10.33851/JMIS.2025.12.1.13

In open world object localization tasks, existing methods place excessive emphasis on learning from high-quality object samples while ne-glecting low-quality samples and background samples, which are considered negative samples. This tendency not only hinders the model’s ability to explore the features of object candidate regions but also limits further improvements in the quality of training samples. To address this issue, we propose an open world object localization method based on negative sample learning. We design a negative sample generation model to produce background samples or low-quality object samples with weaker features, aiming to balance the sample distribution and re-duce the risk of the model learning incorrect features. By incorporating low-quality object samples as negative samples in the localization branch and training them alongside positive samples, we effectively mitigate the issue of sample imbalance and reduce the instability of sam-pling strategies. Meanwhile, the generated background samples are used as negative samples for the classification branch, minimizing interfer-ence from unlabeled object samples in background sampling. Additionally, in open world learning scenarios, the model is prone to interference from unknown class samples during the classification phase. To address this, we introduce an Out-of-Distribution (OOD) detector before the classification task. This detector can efficiently filter out unknown class samples, reducing their negative impact on the model’s classification performance and thereby enhancing the classification accuracy for known categories. To validate the effectiveness of the proposed method, we conducted experiments on the COCO dataset. The results show that our method achieves significant improvements in both localization and classification accuracy compared to other methods. KCI Citation Count: 0