Short Bio

I am a PhD candidate in Electrical Engineering at University of California, Riverside. I work under the supervision of Dr. Bir Bhanu at Visualization and Intelligent Systems Laboratory (VISLab). My research interest lies in Computer Vision, especially in video analysis research such as video recognition (classification, detection, segmentation) and action recognition. I got my M.S. from University of Southern California and B.Sc. from Beijing University of Posts and Telecommunications.

---

Publications

1. Pose-Guided R-CNN for Jersey Number Recognition in Sports.

   By Liu, H. and Bhanu, B.

   In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops2019.
   Abstract BibTeX PDF

   @inproceedings{Liu_2019_CVPR_Workshops,
     author = {Liu, Hengyue and Bhanu, Bir},
     title = {Pose-Guided R-CNN for Jersey Number Recognition in Sports},
     booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
     month = jun,
     year = {2019}
   }
   

   Recognizing player jersey number in sports match video streams is a challenging computer vision task. The human pose and view-point variations displayed in frames lead to many difficulties in recognizing the digits on jerseys. These challenges are addressed here using an approach that exploits human body part cues with a Region-based Convolutional Neural Network (R-CNN) variant for digit level localization and classification. The paper first adopts the Region Proposal Network (RPN) to perform anchor classification and bounding-box regression over three classes: background, person and digit. The person and digit proposals are geometrically related and fed to a network classifier. Subsequently, it introduces a human body key-point prediction branch and a pose-guided regressor to get better bounding-box offsets for generating digit proposals. A novel dataset of soccer-match video frames with corresponding multi-digit class labels, player and jersey number bounding boxes, and single digit segmentation masks is collected. Our framework outperforms all existing models on jersey number recognition task. This work will be essential to the automation of player identification across multiple sports, and releasing the dataset will ease future research on sports video analysis.

2. Context-Aware Control for Dynamic Execution of Throttleable Neural Networks.

   By Liu, H., Parajuli, S., Hostetler, J., Chai, S. and Bhanu, B.

   2020 (Under review).
   Abstract BibTeX

   @inproceedings{Liu_2020_conf,
     title = {Context-Aware Control for Dynamic Execution of Throttleable Neural Networks},
     author = {Liu, Hengyue and Parajuli, Samyak and Hostetler, Jesse and Chai, Sek and Bhanu, Bir},
     year = {2020 (Under review)}
   }
   

   Conditional computation approaches have been proposed to reduce overall computational load and improve model accuracy by selecting only a small portion of the neural model for processing. This paper presents a runtime-throttleable neural network (TNN) that can adaptively regulate its own performance target and computing resources. TNNs are composed of modular blocks that can be gated with a single input-dependent utilization parameter. A context-aware controller is trained separately to directly optimize arbitrary application-level performance metrics. Extensive experiments on CIFAR-10, ImageNet and VOC2007 datasets show that TNNs can be effectively throttled across a range of utilization settings, while having peak accuracy comparable to the vanilla networks. We present results for a context-aware TNN-based gesture recognition system that demonstrates the effectiveness of throttling using spatio-temporal features over a set of operating conditions and accuracy levels.

---