Kong Johnny

Kong Johnny

Novice scholar in deep learning, computer vision and 3D reconstruction.

Posts by Collection

portfolio

publications

Enhanced Inpainting Model Revitalizes Historical Paintings with Vision Transformer

Published in ICVR 2023, 2023

This paper presents a deep learning architecture for restoring ancient paintings, which have immense historical and artistic value as they vividly record history from diverse perspectives. Due to the passage of time, many historical works have suffered damage, which requires time-consuming manual restoration by skilled professionals. Our proposed method utilizes a sophisticated edge detection model to extract structure information from the paintings, including texture, painting style, and structure, which are applied for restoration. The effectiveness of the proposed method was validated by training and testing on various ancient painting datasets. This work has significant value in that it can expedite and enhance the accuracy of the restoration process without compromising the original artistic style and intent, thereby better preserving and transmitting our historical culture. We believe that the contribution of this work is meaningful for VR cultural heritage conservation and presentation.

Recommended citation: X. Duan, C. Jiang and Y. Fan, "Enhanced Inpainting Model Revitalizes Historical Paintings with Vision Transformer," 2023 9th International Conference on Virtual Reality (ICVR), Xianyang, China, 2023, pp. 582-589, doi: 10.1109/ICVR57957.2023.10169621. http://kong-johnny.github.io/files/paper1.pdf

ST-4DGS: Spatial-Temporally Consistent 4D Gaussian Splatting for Efficient Dynamic Scene Rendering

Published in SIGGRAPH 2024, 2024

Dynamic scene rendering at any novel views continues to be a difficult but important task, especially for high-fidelity rendering quality at real-time rendering speed. The recent 3D Gaussian Splatting, i.e., 3DGS, shows great success for static scene rendering with impressive quality at very efficient speed, however, the extension of 3DGS from static scene to dynamic as 4DGS is still challenging, especially to maintain the spatial-temporally persistent dynamic rendering quality. This paper proposes a novel spatial-temporally consistent 4D Gaussian Splatting, i.e., ST-4DGS, for real-time dynamic scene rendering which especially aims at the spatial-temporally persistent dynamic rendering quality and maintain the real-time efficiency. The key ideas of ST-4DGS are two new novel mechanisms: (1) a spatial-temporal encoder for 4D Gaussians representations but with a motion-aware shape regularization, and (2) a spatial-temporally joint density control, which are very effective to prevent the compactness degeneration for the 4D Gaussians representation during the dynamic scene learning, thus leading to spatial-temporally consistent dynamic rendering quality while in efficient manner. With extensive evaluation on public datasets, our ST-4DGS can achieve much better dynamic rendering quality than previous dynamic rendering approaches, such as 4DGS, HexPlane, K-Plane etc, and in a more efficient way for real-time dynamic rendering. To our best knowledge, we are the first 4D Gaussian Splatting for high-fidelity dynamic rendering at real-time speed, especially ensuring the spatial-temporally consistent rendering quality.

Download here

talks

teaching

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.

Undergraduate educational experience

Undergraduate course, Beijing Normal University, School of Artificial Intelligence, 2020

This is a description of a teaching experience. You can use markdown like any other post.