LinYark

Hi, there, I'm LinYark. My real name is Fengchen He, 何丰郴.
I am an Optical Engineering Ph.D. student in HUST, working with Prof. Shaoqun Zeng.

My research interest is to simulate the imaging process of cameras to correct the downstream CV task.
Dual-Pixel cameras are currently the focus of my research.
Our team is exploring a variety of implementations centered on camera simulation and downstream applications, including dual-pixel, bokeh, VLM.

Passionate about the fields of programming, swimming, electronics, and photography. Crossed the Yangtze River multiple times.

Email  /  Github  /  知乎  /  Wechat

profile photo
News

  • 07/2026: "FoundDP: Revisiting Weak Disparity Observability in Dual-Pixel Depth Estimation" is accepted by ECCV 2026.
  • 07/2025: "Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation" is accepted by ICCV 2025.
  • 03/2025: I have shared a ray-traced Dual-Pixel simulator detailing how to simulate realistic Dual-Pixel images.
  • 07/2024: I maintain a repository dedicated to awesome Dual-Pixel camera technology.
  • 09/2023: I commenced my doctoral journey at HUST. Advisor Shaoqun Zeng.
  • 03/2022: I worked as an Auto-Focus algorithm engineer at Xiaomi. PM Yongfu Song. Miss all my dear colleagues.

    		•
    Education & Working

  • 09/2023 - Now: Ph.D. in Optical Engineering, HUST, Wuhan, China.
  • 03/2022 - 09/2023: Algorithm Engineer on Auto-Focus, Camera Department, Xiaomi, Shanghai, China.
  • 09/2019 - 03/2022: M.Sc. in Optical Engineering, NJUST, Nanjing, China.
  • 09/2015 - 06/2019: B.Sc. in Optical Engineering, NJUST, Nanjing, China.

    		•
    Research & Project

    I am conducting simulations of cameras and then applying them to various applications.

  • Computational Dual-Pixel cameras.
  • Optical imaging system.
  • Depth est, deblur, bokeh and so on.

    		•
    		OABokeh: Occlusion-Aware Bokeh Rendering from 2.5D RGB
    Fengchen He, Xinge Yang, Hao Xu, Dayang Zhao, Shaoqun Zeng

  • Models bokeh as visibility-modulated aperture integration from single 2.5D RGB.
  • Generates occlusion-truncated spatial-variant blur kernels to eliminate boundary artifacts.
  • Lightweight CNN acceleration achieves 25× speedup with slight fidelity loss.

    		•
    		FoundDP: Revisiting Weak Disparity Observability in Dual-Pixel Depth Estimation
    Fengchen He, Hao Xu, Dayang Zhao, Tingwei Quan, Shaoqun Zeng
    ECCV 2026. Paper, Github, Website

  • Integrates DP metric depth with foundation-model structural priors.
  • Aligns degraded ViT features for stable depth guidance.
  • Improves structure and accuracy in weak-disparity regions.

    		•
    		Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation
    Fengchen He, Dayang Zhao, Hao Xu, Tingwei Quan, Shaoqun Zeng
    ICCV 2025. Paper , Github, Website

  • The ray-traced DP PSF simulator calculates the spatially varying DP PSF.
  • The pixel-wise DP image rendering module outputs realistic DP simulated images.
  • The depth estimation model trained with Sdirt generalizes better.
  • Released a real test set DP119, with an open lens structure and fixed focus.

    		•
    		Auto-Focus Algorithm
    PM Yongfu Song, PL Xilei Cai, Mentor Yanhui Wang. Miss U.

  • Phase, contrast, ToF, and AI autofocus.
  • The CamX part of autofocus.
  • Participated in: Xiaomi Mix Fold 2 and 3, Redmi K60 Pro and K70 Pro.

    		•
    		SiamDL: Siamese Dual-Level Fusion Attention Network for RGBT Tracking
    Fengchen He, Mingyang Chen, Xiaoyu Chen, Jing Han, Lianfa Bai
    Paper, Github

  • Cross-domain siamese attention updates templates with rich mode and time context.
  • Dual-level balance module fuses feature and decision data to balance modal weights.
  • Our SiamBAN-based method adds MDAM and DLBM, achieving competitive tracking results.

    		•
    		Autonomous Remote Strike System
    Based on Linux and Qt. Not a paper, just a project in 2021s.
    System overview:

  • UAV equipped with TX2 and Realsense camera communicates via serial protocol.
  • At high altitude, detection and tracking algorithms lock the target object.
  • At low altitude, secondary detection extracts detailed target parts for tracking.
  • Depth maps from Realsense correct positioning, updating flight control.

    		•
    		Automated Welding System
    Based on Linux and Qt. Not a paper, just a project in 2020s.
    System overview:

  • Parallel cameras extract 2D workpiece features via segmentation and edge detection.
  • Hand-eye calibration maps 2D features to robot coordinates, forming 3D point clouds.
  • OCR extracts workpiece names to retrieve matching real point clouds from the database.
  • Registration aligns real and captured clouds to locate accurate welding positions.

    		•
    Miscellanea

  • Peer review statement: I am honored to be invited to serve as a peer reviewer in the Dual-Pixel camera field. I will make every effort to complete the review within two weeks.

    		•

    The website template is from Dr. John Barron and Dr. Xinge Yang.