Hi there 👋

My name is Shengli Zhou, and I am a postgraduate student majoring in Computer Science at The Chinese University of Hong Kong (CUHK). Prior to my master’s studies, I completed my undergraduate program at Southern University of Science and Technology (SUSTech), where I was awarded an Honors Bachelor of Engineering degree. I am actively seeking full-time job opportunities commencing in Fall 2027.

🔭 My personal website: fz-zsl.github.io
🌱 I’m currently working on multimodal learning for 3D vision-language tasks and short drama generation.
📫 How to reach me: zhousl2004@outlook.com
🔥 Enture forth to embrace the boundless unknown (去看看未见的广袤！)

Education

Postgraduate: Sep. 2026 - Jul. 2027 (Expected)

The Chinese University of Hong Kong (CUHK)

Department of Computer Science and Engineering (CSE)

Program: Master of Science in Computer Science

Undergraduate: Sep. 2022 - Jul. 2026

Southern University of Science and Technology (SUSTech)

Department of Computer Science and Engineering (CSE) & Zhiren College

Major: Computer Science and Technology (GPA: 3.98 / 4.00, rank: 1 / 167)

Member of Turing Class (designated for elite CS students at SUSTech)

Selected Publications

Scalable Object Relation Encoding for Better 3D Spatial Reasoning in Large Language Models

Shengli Zhou, Minghang Zheng, Feng Zheng, Yang Liu✉
Accepted by CVPR 2026 Main Conference: Project Page / Paper (CVF) / Paper (arXiv) / Code / 机器之心
Presentation: CVPR Poster Page / Poster / Video / Slides
This paper proposes QuatRoPE, a linear-scalable 3D positional embedding method that computes pairwise object spatial relations via quaternion rotations in Transformer attention layers, and the Isolated Gated RoPE Extension (IGRE) to minimize its interference with LLMs’ original language RoPE, while also introducing the ASR benchmark for pure 3D spatial reasoning evaluation; extensive experiments show that the proposed methods consistently boost the 3D spatial reasoning performance of LLMs on multiple 3D VL benchmarks and the ASR benchmark, outperforming strong baselines and validating their effectiveness.

CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models

Shengli Zhou, Xiangchen Wang, Guanhua Chen✉, and Feng Zheng✉
Accepted by ACL 2026 Main Conference: Project Page / Paper (ACL Anthology) / Paper (arXiv) / Code
Presentation: ACL Conference Page & Video / Poster / Slides
Existing scene graph pruning for 3D vision-language tasks often discards task-critical relations, harming spatial reasoning. To address this issue, we propose CAPruner, which combines semantic relevance and spatial proximity to estimate relation importance under specific task context, trained without expensive relation-level annotations. Experiments show it preserves key spatial relations and significantly boosts LLM performance on 3D-VL tasks.

Learn 3D VQA Better with Active Selection and Reannotation

Shengli Zhou, Yang Liu, Feng Zheng✉
Accepted by ACM MM 2025: Paper (ACM Digital Library) / Paper (arXiv) / Code / 公众号
To address the negative impact of inevitable improper annotation in 3D Visual Question-Answering and the scarcity of annotations, we propose a multi-turn interactive active learning strategy, combining semantic variance-based data selection with interactive oracle reannotation, enhancing answer quality and reducing training costs.

Click here to see the full list.

Research & Visiting Experience

Feb. 2025 - Jul. 2025

Peking University (PKU)

Wangxuan Institute of Computer Technology (WICT)

Multimedia Information Processing Lab (MIPL)

Supervisor: Prof. Yang Liu

May 2024 - Jul. 2024

National University of Singapore (NUS)

School of Computing (SoC) Summer Workshop (SWS) 2024

Visual Computing | Lecturer: Prof. Terence Sim | Poster

Selected Projects

When Active Learning and Data Augmentation Meet at Object Detection

Autonomous driving systems face challenges in real-time object detection due to high labeling costs and limited dataset diversity, prompting optimization of the RT-DETR-v2 model.
Integrated active learning (AL) and data augmentation to enhance detection accuracy while minimizing labeling efforts, focusing on evaluating AL strategies (random, entropy-based, information gain) and augmentation levels (e.g., motion blur, noise).
Achieved 75.4% mAP (6% improvement) on KITTI, validating entropy-based AL as optimal and medium-strength augmentation for balancing accuracy and real-world alignment, enabling efficient data utilization. Github

Masked-Unmasked Face Recognition

Face recognition struggles with brightness, angle, expression, and occlusion variations.
This project uses a non-deep learning approach to develop a classifier inspired by human recognition. It focuses on selecting high-quality local features to improve accuracy and reliability in face recognition. The classifier adaptively selects features based on current conditions, excluding less relevant local features to enhance recognition accuracy.
A mask detector with 100% accuracy on the Georgia Tech Dataset and maintaining high accuracy on harder datasets. Poster