Unsupervised detector adaptation by joint dataset feature learning

Object detection is an important step in automated scene understanding. Training state-of-the-art object detectors typically require manual annotation of training data which can be labor-intensive. In this paper, we propose a novel algorithm to automatically adapt a pedestrian detector trained on a generic image dataset to a video in an unsupervised way using joint dataset deep feature learning. Our approach does not require any background subtraction or tracking in the video. Experiments on two challenging video datasets show that our algorithm is effective and outperforms the state-of-the-art approach.