Recent Few-shot Object Detection Algorithms: A Survey with Performance Comparison

• DOI: 10.1145/3593588
• Publication Date: 2023-05-02T12:39:19Z
• Authors: Tianying Liu; Lu Zhang; Yang Wang; Jihong Guan; Yanwei Fu; Jiajia Zhao; Shuigeng Zhou
• Keywords: Transfer of learning

The generic object detection (GOD) task has been successfully tackled by recent deep neural networks, trained an avalanche of annotated training samples from some common classes. However, it is still non-trivial to generalize these detectors the novel long-tailed classes, which have only few labeled samples. To this end, Few-Shot Object Detection (FSOD) topical recently, as mimics humans’ ability learning learn and intelligently transfers learned knowledge heavy-tailed Especially, research in emerging field flourishing years with various benchmarks, backbones, methodologies proposed. review FSOD works, there are several insightful survey articles [ 58 , 59 74 78 ] that systematically study compare them groups fine-tuning/transfer meta-learning methods. In contrast, we existing algorithms a new perspective under taxonomy based on their contributions, i.e., data-oriented, model-oriented, algorithm-oriented. Thus, comprehensive performance comparison conducted achievements FSOD. Furthermore, also analyze technical challenges, merits demerits methods, envision future directions Specifically, give overview FSOD, including problem definition, datasets, evaluation protocols. then proposed methods into three types. Following taxonomy, provide systematic advances Finally, further discussions performance, presented.