Biview learning framework for human posture segmentation.

<p>The first row demonstrates extracted original high-dimensional depth features (DDF), and then using the training data, we apply DLA as our first stage for dimension reduction to obtain more discriminative features. From the training sample, apparently, the points on different body parts are separated with high margins. The second row demonstrates the extracted unsupervised relative position features. By CCA, it tries to explore complement information, namely, using the unsupervised RPF adjusts the overfitting of the learned features while using learned DDF features to introduce more discriminative ability. Finally, the k-d features (k is much less than the dimensions of DDF) are inputted to train a traditional SVM classifier.