Wang Xin ; Huang Thomas E. ; Darrell Trevor ; Gonzalez Joseph E. ; Yu Fisher

Abstract

Detecting rare objects from a few examples is an emerging problem. Priorworks show meta-learning is a promising approach. But, fine-tuning techniqueshave drawn scant attention. We find that fine-tuning only the last layer ofexisting detectors on rare classes is crucial to the few-shot object detectiontask. Such a simple approach outperforms the meta-learning methods by roughly2~20 points on current benchmarks and sometimes even doubles the accuracy ofthe prior methods. However, the high variance in the few samples often leads tothe unreliability of existing benchmarks. We revise the evaluation protocols bysampling multiple groups of training examples to obtain stable comparisons andbuild new benchmarks based on three datasets: PASCAL VOC, COCO and LVIS. Again,our fine-tuning approach establishes a new state of the art on the revisedbenchmarks. The code as well as the pretrained models are available athttps://github.com/ucbdrive/few-shot-object-detection.