Abstract

This paper studies the challenging cross-domain few-shot object detection(CD-FSOD), aiming to develop an accurate object detector for novel domains withminimal labeled examples. While transformer-based open-set detectors, such asDE-ViT, show promise in traditional few-shot object detection, theirgeneralization to CD-FSOD remains unclear: 1) can such open-set detectionmethods easily generalize to CD-FSOD? 2) If not, how can models be enhancedwhen facing huge domain gaps? To answer the first question, we employ measuresincluding style, inter-class variance (ICV), and indefinable boundaries (IB) tounderstand the domain gap. Based on these measures, we establish a newbenchmark named CD-FSOD to evaluate object detection methods, revealing thatmost of the current approaches fail to generalize across domains. Technically,we observe that the performance decline is associated with our proposedmeasures: style, ICV, and IB. Consequently, we propose several novel modules toaddress these issues. First, the learnable instance features align initialfixed instances with target categories, enhancing feature distinctiveness.Second, the instance reweighting module assigns higher importance tohigh-quality instances with slight IB. Third, the domain prompter encouragesfeatures resilient to different styles by synthesizing imaginary domainswithout altering semantic contents. These techniques collectively contribute tothe development of the Cross-Domain Vision Transformer for CD-FSOD (CD-ViTO),significantly improving upon the base DE-ViT. Experimental results validate theefficacy of our model.