Transferred IR pedestrian detector toward distinct scenarios adaptation

The presence of inevitable disparity in distributions between the training data and test data is one of the main reasons that result in downgraded performance for infrared (IR) pedestrian detection across distinct scenarios, and it is expensive and sometimes difficult to label sufficient new training data from target scenarios to re-train a scene-specific detector. In this paper, a novel boosting-style method for data-level transfer learning termed DTLBoost is proposed. Specifically, sample importance measurement is first presented to evaluate the similarities between the samples across distinct scenarios using k-nearest neighbors-based model. Then the most informative samples from former scenarios are selected to extend the training data and help to build a base learner iteratively. In addition, degree of classification disagreement among base learners is formulated and incorporated into the weight updating rules of training samples, which helps to select the samples from former scenarios with positive transferability and encourage different base learners to learn different parts or aspects of samples from target scenarios. The proposed method has been evaluated on two types of IR pedestrian detection applications, including pedestrian detection for both driving assistance systems and video surveillance. Experimental results demonstrate that the proposed method achieves promising improvement on detection performance toward both new scenes and viewpoints adaptation.