HOME RESEARCH C.V.

Pyramid Structure Optical Flow Learning with Motion Cue

Ji Dai, Shiyuan Huang, Truong Nguyen

2018 IEEE International Conference on Image Processing, Athens, Greek

Abstract: Presents a pyramid structured framework for optical flow estimation. At each pyramid scale, a shallow network predicts an update optical flow from previous scale's prediction. Proposed method reaches 100 fps on flying chair dataset with state-of-the-art accuracy.(oral presentation) [pdf]

Accurate and Efficient Video De-fencing Using Convolutional Neural Networks and Temporal Information

Chen Du, Byeongkeun Kang, Zheng Xu, Ji Dai, Truong Nguyen

2018 IEEE International Conference on Multimedia and Expo, San Diego, CA.

Abstract: We present a learning method for fense detection, which significantly improve the accuracy over state-of-the-art. We also release the large scale fense dataset with precise ground truth annotation used to train the network.[pdf]

View Synthesis with Hierarchical Clustering Based Hole Filling

Ji Dai, Truong Nguyen

2017 IEEE International Conference on Image Processing, Beijing, China

Abstract: Presents a depth image based rendering algorithm for view synthesis. Challenging occlusion filling problem is addressed with a hierarchical clustering based restoration method. Proposed algorithm achieves best scores on middlebury dataset. (oral presentation) [pdf] [code]

Towards Privacy-Preserving Activity Recognition Using Extremely Low Temporal and Spatial Resolution Cameras

Ji Dai, Jonathan Wu, Behrouz Saghafi, Janusz Konrad, Prakash Ishwar

2015 Workshop on Analysis and Modeling of Faces and Gestures (AMFG) in Conjuction with CVPR, Boston, MA

Abstract: In this paper, we generally improve the performance of the recognition algorithm in previous paper. We include Temporally warping in the feature descriptor and extend the evaluation from synthetic data to real data (IXMAS). (oral presentation) [pdf]

Towards Privacy-Preserving Recognition of Human Activities

Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, Prakash Ishwar

2015 IEEE International Conference on Image Processing, Quebec City, Canada

Abstract: This paper addresses the concern of jeopardizing privacy when applying activity recognition devices to everyday environments. Instead of using high resolution cameras, we propose an recognition algorithm that works with extremely low resolution cameras (< 10x10). [pdf]