Research

Select Projects

Temporal Context Network for Activity Localization in Videos     (ICCV 2017)   [Code]

• • A Faster-RCNN like proposal and detection architecture

  • Proposal anchors are placed at equal intervals in a video which span multiple temporal scales

  • Pair-wise input sampling to capture context

  • Temporal convolution is used to learn temporal consistency and position of boundaries

  • Top 3 method in Activity-Net 2017 Challenge for Action Proposals

  • We demonstrate that building a representation by sampling from a pair of scales which explicitly captures context around proposal anchors is important for precise temporal localization

Efficient Fine-grained Classification and Part Localization Using One Compact Network       (ICCVW 2017)

• • We propose a novel multi-task deep learning and fusion architecture that have both shared and dedicated convolutional layers for simultaneous part labeling and make-model classification. 

  • The accuracy of our approach is competitive to state-of-the-art methods on both car and bird domains.

  • Our network architecture is more compact (30M parameters) and runs much faster (78FPS) than competitors, enabling real-time, mobile applications. 

FASON: First and Second Order Information Fusion Network     (CVPR 2017)

• • We design a deep fusion architecture that effectively combines second order information (calculated from a bilinear model) and first order information (preserved through our leaking shortcut) in an end-to-end deep network. 

  • Our architecture enables more effective learning.

  • We extend our fusion architecture to take advantage of the multiple features from different convolution layers.