Boyu Cai

I am a joint M.S. student in Computer Science at ShanghaiTech University and the Institute of Automation, Chinese Academy of Sciences, advised by Weiming Hu, Li Yang, and Yan Xu.

My research focuses on learning generalizable visual representations for understanding open and dynamic environments. Across my work, I have found that language consistently provides semantic structure that strengthens grounding and generalization.

I am currently open to Ph.D. opportunities and industry positions. Please feel free to contact me!

Research

Across these projects, two lessons have shaped how I approach research: (1) end-to-end training often coordinates the full system more effectively than fine-tuning modules in isolation; and (2) well-designed model priors provide crucial structure for learning and generalization.

SPAR teaser comparing dynamic-scene reconstruction and training paradigms

Dynamic-Robust Photometric-Semantic Reconstruction for Open-Vocabulary 3D Scene Understanding

Boyu Cai, Li Yang, Yan Xu, Wei Liu, Nian Liu, Sikui Zhang, Yan Wang, Chunfeng Yuan, Weiming Hu

ECCV, 2026

project page / arXiv / code

SPAR jointly reconstructs photometric and open-vocabulary semantic scene content from sparse, unposed, and dynamic observations without explicit 3D supervision or ground-truth motion masks.

Overview of the SRA-Det framework for semantic-retrieval-augmented open-vocabulary detection

SRA-Det: Learning Omni-Grained Open-Vocabulary Detection Beyond Category Names

Li Yang^*, Boyu Cai^*, Wei Liu, Yan Wang, Chunfeng Yuan, Bing Li, Weiming Hu

^* Equal contribution.

CVPR, 2026

paper / pdf

SRA-Det retrieves multiple semantic facets from token-level text features and uses attribute-augmented data to improve fine-grained open-vocabulary detection.

Overview of the Temporal Representation Enhancement framework for spiking neural networks

Temporal Representation Enhancement: Learning to Forget Dominant Patterns for Enhanced Temporal Spiking Features

Wei Liu, Li Yang, Yufei Wang, Han Xiao, Boyu Cai, Weiming Hu

CVPR, 2026

project page / pdf

TRE selectively suppresses dominant temporal patterns so spiking neural networks learn more diverse and complementary features across timesteps.

Overview of the Guiding Multi-Task Training framework for end-to-end person search

Focus More on What? Guiding Multi-Task Training for End-to-End Person Search

Boyu Cai, Huibing Wang, Mingze Yao, Xianping Fu

IEEE Transactions on Circuits and Systems for Video Technology, 2025

paper / doi

A guiding multi-task framework addresses imbalanced and conflicting optimization between detection and re-identification in end-to-end person search.

Overview of the Temporal-Self-Erasing supervision framework for spiking neural networks

Towards More Discriminative Feature Learning in SNNs with Temporal-Self-Erasing Supervision

Wei Liu, Li Yang, Mingxuan Zhao, Dengfeng Xue, Shuxun Wang, Boyu Cai, Jin Gao, Wenjuan Li, Bing Li, Weiming Hu

AAAI, 2025

paper / doi

Temporal-Self-Erasing supervision dynamically suppresses previously activated regions, encouraging complementary and more discriminative features over time.

Experience

CyboPal

Apr. 2026 – Jul. 2026

Research Intern

Studied instruction following and out-of-distribution generalization in vision-language-action (VLA) models, using language-anchored action queries to strengthen semantic grounding.
Built a dual-arm real-robot evaluation benchmark and supported robot-client testing and teleoperation data-collection pipelines.

Institute of Automation, Chinese Academy of Sciences (CASIA)

Dec. 2023 – Sep. 2024

Research Intern

Conducted open-vocabulary detection research in a CASIA–Alibaba collaboration, including data construction, model development, and training.
Adapted models to MMDetection and explored hard-negative category sampling to improve open-vocabulary recognition.

Projects

DDSS system architecture: an SDN controller manages dynamic scrubbing policies, P4 programmable switches, and traffic-scrubbing virtual network functions — System architecture · Fig. 4
中文报告 / English report

DDSS: Dynamic DDoS Traffic Scrubbing System

P4-DDOS · Sep. 2021 – Jun. 2023

code report (中文) report (English) project details (English) architecture

A dynamic traffic-scrubbing system combining P4-programmable switches with virtual network functions. An SDN controller periodically updates scrubbing rules using moving-target defense, with game-theoretic analysis guiding policy selection. Implemented and evaluated with P4Runtime, BMv2, Mininet, and DPDK-based scrubbing functions.