Qilong Kou

ORCID: 0000-0002-5222-7069
Publications
Citations
Views
---
Saved
---
About
Contact & Profiles
Research Areas
  • Computer Graphics and Visualization Techniques
  • Human Pose and Action Recognition
  • Advanced Vision and Imaging
  • Human Motion and Animation
  • Video Analysis and Summarization
  • 3D Shape Modeling and Analysis
  • Image and Object Detection Techniques
  • Generative Adversarial Networks and Image Synthesis
  • 3D Surveying and Cultural Heritage
  • Image Enhancement Techniques
  • Industrial Vision Systems and Defect Detection
  • Metaheuristic Optimization Algorithms Research
  • Advanced Numerical Analysis Techniques
  • Virtual Reality Applications and Impacts
  • Power Line Inspection Robots
  • Digital Games and Media
  • Advanced Multi-Objective Optimization Algorithms
  • Video Coding and Compression Technologies
  • Image Processing and 3D Reconstruction
  • Advanced Neural Network Applications
  • Artificial Immune Systems Applications
  • Photopolymerization techniques and applications
  • Vehicle License Plate Recognition
  • Visual Attention and Saliency Detection

Tencent (China)
2022-2024

Real-time in-between motion generation is universally required in games and highly desirable existing animation pipelines. Its core challenge lies the need to satisfy three critical conditions simultaneously: quality, controllability speed , which renders any methods that offline computation (or post-processing) or cannot incorporate (often unpredictable) user control undesirable. To this end, we propose a new real-time transition method address aforementioned challenges. Our approach...

10.1145/3528223.3530090 article EN ACM Transactions on Graphics 2022-07-01

Styled online in-between motion generation has important application scenarios in computer animation and games. Its core challenge lies the need to satisfy four critical requirements simultaneously: speed, quality, style diversity, synthesis controllability. While first two challenges demand a delicate balance between simple fast models learning capacity for latter are rarely investigated together existing methods, which largely focus on either control without or uncontrolled stylized...

10.1145/3588432.3591514 preprint EN 2023-07-19

Recent advancements in 2D diffusion models allow appearance generation on untextured raw meshes. These methods create RGB textures by distilling a model, which often contains unwanted baked-in shading effects and results unrealistic rendering the downstream applications. Generating Physically Based Rendering (PBR) materials instead of just would be promising solution. However, directly PBR material parameters from still suffers incorrect decomposition, such as albedo. We introduce DreamMat ,...

10.1145/3658170 article EN ACM Transactions on Graphics 2024-07-19

Abstract We present a novel learning method using two-stream network to predict cloth deformation for skeleton-based characters. The characters processed in our approach are not limited humans, and can be other targets with representations such as fish or pets. use architecture which consists of mesh-based residual networks learn the coarse features wrinkle forming overall from template mesh. Our may used loose tight-fitting clothing. memory footprint is low, thereby resulting reduced...

10.1007/s41095-023-0344-6 article EN cc-by Computational Visual Media 2024-04-19

This article proposes a cuckoo algorithm (GFCS) based on the global feedback strategy and innovatively introduces "re-fly" mechanism. In GFCS, process of is adjusted controlled by dynamic variable, parameter also serves as an indicator whether has fallen into local optimum. According to change optimum value in each round, variable optimize algorithm. addition, we set new formulas for other main parameters, which are progresses. When converges prematurely falls optimum, current retained,...

10.1155/2023/2040866 article EN cc-by Computational Intelligence and Neuroscience 2023-01-01

We present a novel locality-based learning method for cleaning and solving optical motion capture data. Given noisy marker data, we propose new heterogeneous graph neural network which treats markers joints as different types of nodes, uses convolution operations to extract the local features transform them clean motions. To deal with anomaly (e.g. occluded or big tracking errors), key insight is that marker's shows strong correlations motions its immediate neighboring but less so other...

10.1145/3610548.3618148 preprint EN 2023-12-10

In order to realize the real-time and efficient detection of substation switching device status, solve some existing problems inspection system, such as limited degree automation need for human intervention problem. This paper proposes a status recognition method based on deep learning. The obtains image data by optical camera an robot, uses learning technology analyze detect data. this method, firstly, is selected marked set model training, then Yolov3 target network used build automatic...

10.1109/docs55193.2022.9967743 article EN 2022-10-28

2D diffusion model, which often contains unwanted baked-in shading effects and results in unrealistic rendering the downstream applications. Generating Physically Based Rendering (PBR) materials instead of just RGB textures would be a promising solution. However, directly distilling PBR material parameters from models still suffers incorrect decomposition, such as albedo. We introduce DreamMat, an innovative approach to resolve aforementioned problem, generate high-quality text descriptions....

10.48550/arxiv.2405.17176 preprint EN arXiv (Cornell University) 2024-05-27

Motion style transfer changes the of a motion while retaining its content and is useful in computer animations games. Contact an essential component that should be controlled explicitly order to express vividly enhancing naturalness quality. However, it unknown how decouple control contact achieve fine-grained transfer. In this paper, we present novel method for over contacts achieving both spatial-temporal variations style. Based on our empirical evidence, propose controlling indirectly...

10.1145/3680528.3687609 preprint EN 2024-12-03

Abstract We propose a novel ray reordering technique designed to accelerate the tracing process by encoding and sorting rays prior traversal. Our method, called “hierarchy cut code”, involves based on cuts of hierarchical acceleration structure, rather than relying solely spatial coordinates. This approach allows for more effective adaptation resulting in reliable efficient outcome. Furthermore, our research identifies “bounding drift” as major obstacle achieving better effects using longer...

10.1111/cgf.15226 article EN Computer Graphics Forum 2024-10-01

Optical motion capture (MoCap) is the "gold standard" for accurately capturing full-body motions. To make use of raw MoCap point data, system labels points with corresponding body part locations and solves However, data often contains mislabeling, occlusion positional errors, requiring extensive manual correction. alleviate this burden, we introduce RoMo, a learning-based framework robustly labeling solving optical data. In stage, RoMo employs divide-and-conquer strategy to break down...

10.1145/3680528.3687615 preprint EN 2024-12-03

Optimizing the memory footprint of 3D models can have a major impact on user experiences during real-time rendering and streaming visualization, where overhead lies in high-resolution texture data. In this work, we propose robust automatic pipeline to content-aware, lossy compression for atlas. The design our solution two observations: 1) mapping multiple surface patches same region is seamlessly compatible with standard pipeline, requiring no decompression before any usage; 2) image has...

10.1145/3610548.3618150 article EN 2023-12-10

In order to quickly and accurately detect the display defects of electric meter LCD screen, this paper proposed an screen defect detecting method based on convolutional neural network (CNN). First, a horizontal straight line frame is found by LSD detection for tilt correction. Second, area positioned normalized correlation matching corrected image. Then, accurate positions characters are located position information generated template annotation tool. Finally, CNN used perform character OCR...

10.1109/cac53003.2021.9728544 article EN 2021 China Automation Congress (CAC) 2021-10-22

With the development of intelligent substation construction technology, 3D recognition equipment is becoming more and important. This paper proposes an improved algorithm based on KNN classification subspace feature vector, in which number dividing subspaces unchanged to maximize keeping shape features point cloud, solves problem distorting cloud caused by normalization. At same time, this preselects standard devices from template library size screening reduce comparison range shorten...

10.1109/cac53003.2021.9728083 article EN 2021 China Automation Congress (CAC) 2021-10-22

Abstract Character hit reaction is an inherent component in game development. Natural reactions games are typically achieved through the use of artist‐created animations and motion capture. To improve realism impact reactions, developers combine physics simulation with distinct based on character statuses. However, there currently no method that can automatically produce information this end, we propose a physics‐driven inverse kinematic for generating animations. We postulate character's...

10.1002/cav.2170 article EN Computer Animation and Virtual Worlds 2023-05-01

We propose a novel ray reordering technique to accelerate the tracing process by encoding and sorting rays prior traversal. Instead of spatial coordinates, our method encodes according cuts hierarchical acceleration structure, which is called hierarchy cut code. This approach can better adapt structure obtain more reliable result. also compression scheme decrease overhead shorter key. In addition, based on phenomenon boundary drift, we theoretically explain reason why existing methods cannot...

10.48550/arxiv.2305.16652 preprint EN cc-by arXiv (Cornell University) 2023-01-01

We present a novel learning method to predict the cloth deformation for skeleton-based characters with two-stream network. The processed in our approach are not limited humans, and can be other skeletal-based representations of non-human targets such as fish or pets. use network architecture which consists mesh-based residual networks learn coarse wrinkle features overall from template mesh. Our is used loose tight-fitting clothing dresses. ensure that memory footprint low, thereby result...

10.48550/arxiv.2305.18808 preprint EN other-oa arXiv (Cornell University) 2023-01-01

The path tracing method generates incoherent rays by randomly sampling directions. This randomness makes it unsuitable for modern processor architectures that rely on coherence to achieve optimal performance. Many efforts have been made address this issue reordering based their origin, end, or direction enhance coherence. However, a drawback of methods is the need encode and sort before tracing, introducing additional overhead. We propose technique generate coherent directly reusing...

10.48550/arxiv.2310.07182 preprint EN cc-by arXiv (Cornell University) 2023-01-01
Coming Soon ...