Skip to main content

Gavin Gong

a.k.a. VisualDust

I'm currently a final year B.E. in Computer Science and Technology, at School of Computer and Information Security, Guilin University of Electronic Technology in China. I taught myself computer vision related content during my undergraduate studies. I consist of Python, C#, C++, Java, Kotlin, Latex, dotnet, React, PyTorch, and LaTeX.

My research interests span the broad area of computer vision, especially in dence prediction(i.e. semantic segmentation), point clouds and 3D reconstruction, as well as fine tuning and optimizing deep learning models into software systems. I'm also interested in generative models(i.e. diffusion models), and few shot learning(i.e. AutoEncoders).

My ultimate goal is autonomous driving and autonomous control. I enjoy doing research, experiments, and building highly performant systems via software approaches.

I'm looking for PhD/MS CS and other opportunities. You can find my full CV HERE.

As a CV Learner

As a CV Learner

Self taught learning materials by me and my friends. Focusing on Deep learning technique for Computer Vision.

As a Programmer

As a Programmer

Random notes about developing, programming or argorithm related theory. Also some useless anecdotes.

As a researcher

As a researcher

Random paper reading, where some interesting papers about Computer Vision were recapitulated and analyzed.

Cool Things

  • Network based inference

    The inference results and images of the assisted driving system are distributed through the network and are compatible with various platforms through web technology. This allows assisted driving edge computing devices to be deployed outside of existing vehicle-mounted terminals. Use a tablet instead when you don't have a screen on car.

  • Vehicle Distance Detection

    Vehicle distance detection while driving built from target detection and depth estimation ensures driving security. The composite model allows inputs from both monodepth estimation models and phisical sensors such as lidar. The model is optimized for edge computing devices with fast inference speed with our visualization UI support.

  • Blind Spot Visualization

    Visualization on blind spots keeps the surroundings safe. The proposed approach allows lidar data and input from ultrasonic sensors. Both hardware and software are implemented from scratch. The system read sensor values within a optimized latency to to ensure real-time speed. We tried a more intuitive visualization method and corresponding UI design.

  • Illegal Driving Behavior Detection

    A set of integration for detecting illegal driving behaviors from edge computing devices, cloud services to mobile apps developed by us. The driver's facial key point features are used to judge dangerous driving behavior. Results are recorded on server and delivered to user's phone app that users can correct their behaviors on their own.

  • Road scene seg on embedded device

    Light weight yet accurate semantic segmentation model designed and trained for road scence. The proposed model is applied for scene understanding and drivable area detection. The proposed model achieved smooth floating-point precision inference on edge computing device NVIDIA Jetson NX with the speed of 25FPS.

  • Network based inference

    The inference results and images of the assisted driving system are distributed through the network and are compatible with various platforms through web technology. This allows assisted driving edge computing devices to be deployed outside of existing vehicle-mounted terminals. Use a tablet instead when you don't have a screen on car.

1 of 6

Gallery