CD-FSOD: A Benchmark for Cross-domain Few-shot Object Detection
Although few-shot object detection (FSOD) has attracted great research attention, no work yet exists that studies FSOD across the different domains seen in real-world scenarios. In this paper, we propose a new study of the cross-domain few-shot object detection (CD-FSOD) benchmark, consisting of image data from a diverse data domain. On the proposed benchmark, we evaluate state-of-art FSOD approaches, and analyze the impact of detection models and pre-training datasets on performance. The results reveal several key findings: (1) the existing FSOD approaches tend to fall, and even underperform the naive fine-tuning model; 2) the pre-training datasets and detection architectures play an important role, and the right choice can boost the performance of the target tasks significantly. Besides, we also analyze the reasons for existing FSOD approaches' failure, and introduce a strong baseline that uses a mutually-beneficial manner to alleviate the overfitting problem. Our approach is remarkably superior to existing approaches by significant margins (%2.3 on average) on the proposed benchmark and also achieves competitive performance on the FSOD benchmark.
READ FULL TEXT