Learning Discriminative Occlusion Feature For Depth Ordering Inference On Monocular Image:
1. A novel sparsity based learning approach is presented to learn the discriminative features, detect the occlusion edges and generate the reliable occlusion map jointly, which does not only supply sufficient cues for depth ordering inference, but also reduce the solution space efficiently.
2. A novel triple descriptor is proposed for determining the semi-local F-G relationship, which can further decrease the solution space and meanwhile avoid the conflict.
Fig. 1. Geting occlusion edge. Fig.2 Semi-local F-G cue. Fig.3 Global Inference.
Fig.4 Example results on Cornell depth-order dataset and NYU2 dataset (input, occlusion edge, groundtruth and our result).
Non-rigid Object Tracking with Superpixel Templates Matching:
Robust Visual Tracking with MultipleWeighted Bags Learning:Our contributions:
1.Exploiting a decomposition of the image frame into superpixels, we formulate the target appearance model as a multiple instance learning task, which naturally overcomes the limitations caused by the bounding box based modeling approaches.
2. A computational efficient target modeling approach is proposed which is suitable for real-time tracking applications.
3.A two-step process is proposed for confidence assignment in the testing frame, which can guarantee the effectiveness of the confidence assignment.
4.The proposed method outperforms 11 top trackers tested on 38 challenging videos.
Tracking Result (compared with 11 top trackers over 38 challenging videos)
Only 6 trackers with best perfromance are presented.