Abstract:
Active learning aims to reduce the annotation cost by selecting those informative samples to improve training efficiency and network accuracy. However, current active learning methods for object detection have three drawbacks: a) the network architectures of the detector during active learning are fixed without considering its saturation; (b) the detector may fall short in giving credible prediction probabilities on unlabeled data; (c) existing uncertainty measures may lead to homogenization of the samples. To overcome these problems, we propose a novel active learning strategy with dynamic neural architecture adaption for object detection. Specially, we incorporate a neural architecture adaption module that modifies and expands the current detector structure for the variation of incoming data stream. We design several network morphism modifications to enable an efficient and optimal adaption, which avoids the retraining of the detector after changing the network architecture in each round. Furthermore, we introduce Dirichlet calibration to correct the classifier for obtaining credible prediction, and present a clustering sampling scheme to select diverse samples. Experimental results show that the proposed method outperforms the previous state-of-the-art active learning methods with fixed architectures, improving 1.9% mAP on BDD and 1.6% mAP on COCO.
