Chenhongyi Yang

University of Edinburgh chenhongyi.yang@ed.ac.uk

I am a Ph.D. student in the Bayesian and Neural Systems (BayesWatch) group at University of Edinburgh, advised by Dr. Elliot Crowley. My primary research topics are computer vision and deep learning. From May 2020 to Jan. 2021, I worked as a research intern in perception algorithms at TuSimple Inc., supervised by Dr. Naiyan Wang. I got my MSc degree in Computer Science from Boston University in May 2020. During my graduate study, I worked as a research assistant in the Image and Video Computing group, supervised by Prof. Margrit Betke and Dr. Vitaly Ablavsky. I got my BEng degree in Computer Science from the University of Science and Technology of China in 2018.

Research Experience

Non-adversarial Domain Adaptation for Semantic Segmentation

Supervisors: Prof. Margrit Betke

  • Proposed a non-adversarial domain adaptation method for semantic segmentation: BiSIDA.
  • BiSIDA aligns the two domains through the continuous parameterization of them, which is achieved through a bidirectional learning paradigm.
  • The corresponding paper is published on AAAI 2021.

Boston University

Detecting Occluded Objects in Urban Scenes

Supervisors: Prof. Margrit Betke, Dr. Vitaly Ablavsky

  • Proposed a novel bounding box level embedding mechanism: Semantics-Geometry Embedding. The embedding made it possible to determine whether two heavily overlapping boxes belong to the same object.
  • Proposed a new Semantics-Geometry NMS algorithm that was based on the Semantics-Geometry Embedding. It remarkably improved object detection in scenarios with heavy intra-class occlusions.
  • Designed a new object detector: Serial R-FCN. It not only provided the capability to learn the Semantics-Geometry Embedding end-to-end, but also improved object detection accuracy.
  • The proposed method achieved state-of-the-art performance on the task of car detection in the benchmark KITTI dataset and the task of pedestrian detection in the CityPersons dataset by improving the detection recall in heavily-occluded scenes.
  • The corresponding paper is published on ECCV 2020.

Boston University

Reasoning Occlusion Order

Supervisor: Dr. Vitaly Ablavsky

  • Designed the Pairwise Faster R-CNN (PFaster R-CNN), in which the pooled features of overlapping boxes were concatenated and fed into a classification module to predict their depth order.
  • Proposed a rule to construct the depth order matrix for every pair of objects in one image, in which we focused on avoiding redundant computation for the same pair of objects.
  • Helped with generating a synthetic dataset to test the proposed algorithm.

Boston University

Object detection for Satellite Images

Undergraduate Thesis, Supervisor: Prof. Shouhong Wan

  • Reproduced a Faster R-CNN object detector using Tensorflow, and achieved compatible result with the numbers reported in the Faster R-CNN paper on Pascal VOC 2012 dataset.
  • The detector was used to detect different kinds of house using high-resolution satellite images.

USTC

Course Projects

Accelerating AlphaZero-Gomoku with Matrix Computing

CS591: Computational Game Theory

  • Designed a pay-off matrix that can be used by AlphaZero-Gomoku to reduce the number of MCTS, so that the inference speed could be accelerated.
  • PThe Gomoku game was modeled as a two-person zero-sum game, and the payoff matrix for one player was constructed by the average of his winning rate and his opponent's losing rate, which were both predicted by the neural network based on the game state. Then the actual action of the player was produced by computing the best strategy.
  • The accelerated AlphaZero-Gomoku achieved a winning rate of 80% when competing with a pure MCTS agent with 7000 MCTS iterations for one step.

Boston University

Deep 3D Fake Box

CS585: Image and Video Computing

  • Proposed a novel deep learning based algorithm to draw 3D bounding boxes for objects in the 2D images plain called Deep 3D Fake Box.
  • Given the 2D bounding box of an object, the 3D box was constructed by 3 independent parameters that were predicted by a neural network.

Boston University

Minilan Interpreter

Fundamental of Programming Language

  • Developed a Minilan language interpreter that supported variable declaration, arithmetic and bool operations, branch and loop statements for the Minilan programming language.
  • Enabled some advanced function for the interpreter such as garbage collection.

USTC

Publication

Chenhongyi Yang, Vitaly Ablavsky, Kaihong Wang, Qi Feng, Margrit Betke. “Learning to Separate: Detecting Heavily-Occluded Objects in Urban Scenes”, ECCV'2020

Kaihong Wang, Chenhongyi Yang, Margrit Betke. “Consistency Regularization with High-dimensional Non-adversarial Source- guided Perturbation for Unsupervised Domain Adaptation in Segmentation”, AAAI'2021

Chenhongyi Yang, Zehao Huang, Naiyan Wang. "QueryDet: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection", arxiv preprint arXiv:2103.09136, 2021