Object view clasfication by selective transer learning wih GentleBoost

The success of learning a vision based object detection or recognition model is highly dependent on the view information of target examples. Most generic object detection systems attempt to estimate the pose of examples by unsupervised learning method when the prior knowledge is absent. On the other hand, the utilization of existing view information is still limited on some specific tasks. To address this gap, we introduce a selective transfer learning framework in this paper. The boosting based knowledge transfer method, TransferBoost, can borrow existing prior knowledge from other classes of objects in a selective way. Given a well labeled source data set, TransferBoost performs boosting at both the instance level and the task level, transferring knowledge by adjusting the weights of individual instances and source tasks simultaneously. This combination of two levels boosting can select useful knowledge more effectively from a mix of relevant and irrelevant source data. On top of that, we propose an improved version of TransferBoost by replacing the base learning framework from AdaBoost to GentleBoost. Our experimental results show that the GentleBoost based TransferBoost method shows a higher performance than conventional TransferBoost method while providing a simpler way for implementation. We believe the transfer of view information over different classes of objects will significantly decrease the cost of learning object detection model and thus extend the applicability of existing methods.