Training an object detector incrementally has hardly been explored. In this paper, we propose attentive feature distillation which leverages both bottom-up and top-down attentions to mitigate forgetting in incremental detection. Then, we systematically analyze the proposed distillation method in different scenarios across various domains and categories. We find out that, contrary to common understanding, domain gaps have relatively smaller negative impact on incremental detection, while category differences are the major problem. For the difficult cases, where the domain gaps and especially categories differences are large, we propose an adaptive exemplar sampling method to select diverse and informative samples from entire datasets, to further prevent forgetting. Experimental results show that we achieve the state-of-the-art performance in three different scenarios across seven object detection benchmark datasets.
View on arXiv