1-HKUST: Object Detection in ILSVRC 2014

Cewu Lu      Hao Chen      Qifeng Chen          Hei Law           Yao Xiao      Chi-Keung Tang

Abstract

The Imagenet Large Scale Visual Recognition Challenges (ILSVRC) is the one of the more important big data challenges in the world. We participated in the object detection track of ILSVRC 2014 and received the 4th place among the 38 teams. Our system involves a number of novel techniques on localization and recognition. On localization, initial candidate proposals are generated from selective search, and a novel technique on regressing bounding boxes using deep learning is used for better object localization. On recognition, to represent a candidate proposal, we adopted three features in our system, namely, RCNN features, IFV features and DPM features. Given these powerful features, category-specific combination functions are learned to improve object recognition. Furthermore, background priors and object interaction priors are also learned and applied. Finally, our results compared with other teams.

Downloads

"1-HKUST: Object Detection in ILSVRC 2014" Cewu Lu, Hao Chen*, Qifeng Chen*, Hei Law*, Yao Xiao* and Chi-Keung Tang (arXiv), 2014 (*indicates equal contribution)    [Paper]   [Bibtex Citation]   [Slides]