An Adaptation Framework for Head-Pose Classification in Dynamic Multi-view Scenarios

Abstract

Multi-view head-pose estimation in low-resolution, dynamic scenes is difficult due to blurred facial appearance and perspective changes as targets move around freely in the environment. Under these conditions, acquiring sufficient training examples to learn the dynamic relationship between position, face appearance and head-pose can be very expensive. Instead, a transfer learning approach is proposed in this work. Upon learning a weighted-distance function from many examples where the target position is fixed, we adapt these weights to the scenario where target positions are varying. The adaptation framework incorporates reliability of the different face regions for pose estimation under positional variation, by transforming the target appearance to a canonical appearance corresponding to a reference scene location. Experimental results confirm effectiveness of the proposed approach, which outperforms state-of-the-art by 9.5% under relevant conditions. To aid further research on this topic, we also make DPOSE-a dynamic, multi-view head-pose dataset with ground-truth publicly available with this paper. 1 We are mainly interested in determining the absolute 3D head-pan (horizontal rotation) into one of eight classes, each denoting a quantized 45 o (360/8) pan.

No Matter Where You Are: Flexible Graph-Guided Multi-task Learning for Multi-view Head Pose Classification under Target Motion

“…We perform our experiments on the DPOSE dataset [17]. To our knowledge, there are no other databases for benchmarking multi-view head pose classification performance under target motion.…”

“…Few works estimate head pose fusing information from multiple views [14,17,20,23]. A particle filter is combined with a neural network for pan/tilt classification in [20].…”

“…A weighted distance approach for classifying pose under target motion is proposed in [17]. Upon dividing the space into four quadrants, max-margin distance learning is employed to learn a classifier per region-such a rigid space partitioning scheme will not optimally encode the pose-appearance relationship under motion, with arbitrary camera geometry.…”

“…Fewer still are head pose estimation methods that utilize information from multiple surveillance cameras. Employing a single camera view is insufficient for studying people's behavior in large environments and multi-view images have been exploited to achieve robust pose estimation [14,22,17,20,23]. However, methods such as [14,20] estimate pose as a person rotates in place, but is not freely moving around in the environment.…”

“…As the target moves, the face can appear larger/smaller and face parts can become occluded/visible due to the target's relative position with respect to the camera. We investigated the effect of appearance change on pose classification using the DPOSE dataset [17], which comprises synchronously recorded images of moving persons from four camera views, associated target positions and head pose annotations. Upon dividing the DPOSE space into four quadrants Q1-Q4, we trained an SVM classifier with HOG [7] features extracted from the 4-view images corresponding to a particular quadrant.…”