---
year: 2026
page_title: LSVOS 2026
title: Large-scale Video Object Segmentation
event: Workshop in conjunction with ECCV 2026
event_url: https://eccv.ecva.net
date: 13:30 - 15:30, September 9, 2026 (Afternoon)
location: Palissad South, Malmö Arena
Malmö, Sweden
background: images/workshop2026/background_1.jpg
overlay: false
nav_exclude: paperdates
---
## Latest News {#news}
**[🔥 New - Sep 4]** The workshop venue has been confirmed: **Palissad South, Malmö Arena**.
**[🔥 New - Sep 4]** The [workshop schedule](#schedule) has been announced.
**[🔥 New]** We are excited to announce the new [MUMU Challenge](mumu.html); everyone is welcome to participate!
**[🔥 Update]** The workshop date has been confirmed for the afternoon of September 9, 2026.
**[🔥 Update - Aug 11]** The [leaderboard](#leaderboard) has been announced.
**[Update - Jul 27]** The challenge is over. Thank you for your participation!
**[🔥 Update - Jul 10]** The evaluation servers for all three challenge tracks are now available: [MOSEv2](https://www.codabench.org/competitions/17494/) / [MeViSv2-Text](https://www.codabench.org/competitions/17496/) / [MeViSv2-Audio](https://www.codabench.org/competitions/17495/).
**[News]** ~~Workshop paper submissions are open through the [OpenReview submission portal](https://openreview.net/group?id=thecvf.com/ECCV/2026/Workshop/LSVOS).~~ **Submissions are closed. Thank you for your participation!**
## Introduction {#intro}
The 8th LSVOS challenge will be held in conjunction with [ECCV 2026](https://eccv.ecva.net) in Malmö, Sweden.
This year, the challenge has three tracks: one VOS track and two RVOS tracks. In the Video Object Segmentation (VOS) track, we will utilize [LVOS](https://lingyihongfd.github.io/lvos.github.io/) and [MOSE](https://mose.video/) to study VOS under more challenging complex environments. LVOS is designed for long-term videos, dealing with complex object motion and long-term reappearance, while MOSE focuses on complex scenes, covering aspects such as object disappearance and reappearance, inconspicuous small objects, heavy occlusions, and crowded environments.
For the two Referring Video Object Segmentation (RVOS) tracks, we will continue to use [MeViS](https://henghuiding.github.io/MeViS/), with text- and audio-based guidance. MeViS focuses on identifying the target object in a video based on motion-related descriptions rather than static attributes. This approach challenges the foundational design principles of existing RVOS methods and encourages a deeper exploration and reevaluation of motion modeling.
In addition, we will hold a series of talks by leading experts in video understanding and embodied intelligence. The following topics will be covered:
- Semantic and panoptic segmentation for images and videos
- Video object segmentation in complex scenes
- Long-term video object segmentation
- Referring video object segmentation
- Video segmentation with motion expressions
- Vision and language
- Cognitive models of object perception
- Real-world understanding and embodied intelligence
## Workshop Schedule {#schedule}
The workshop will be held on September 9, 2026 (Afternoon) at **Palissad South, Malmö Arena**.
| Local Time | Programme |
| --- | --- |
| 13:30 - 13:40 | Opening Remark |
| 13:40 - 14:15Invited Speaker | Session 1Bio
Amir Bar is an Assistant Professor at Imperial College London and a founding member of AMI Labs. Previously, he was a Research Scientist and Postdoctoral Researcher at FAIR after completing his PhD at Tel Aviv University. Amir is a recipient of the Blavatnik PhD Award, a CVPR 2025 Best Paper Honorable Mention Award, and an ICLR 2026 World Models Workshop Outstanding Paper Award. His work on video models also won the Ego4D CVPR 2022 PNR Challenge. Before beginning his PhD, Amir led the AI team at Zebra Medical Vision, which was later acquired, and developed multiple FDA-approved algorithms that are currently used in clinical practice worldwide.
Bio
Jiajun Wu is an Assistant Professor of Computer Science and, by courtesy, of Psychology at Stanford University, working on computer vision, machine learning, robotics, and computational cognitive science. Before joining Stanford, he was a Visiting Faculty Researcher at Google Research. He received his PhD in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology. Wu's research has been recognized through the IJCAI Computers and Thought Award, the Young Investigator Programs (YIP) by ONR and by AFOSR, the NSF CAREER award, the Okawa research grant, the AI's 10 to Watch by IEEE Intelligent Systems, paper awards and finalists at ICCV, CVPR, SIGGRAPH Asia, ICRA, CoRL, and IROS, dissertation awards from ACM, AAAI, and MIT, the 2020 Samsung AI Researcher of the Year, and faculty research awards from Microsoft, Google, Nvidia, J.P. Morgan, Samsung, Amazon, and Meta.
Yanwei LiShanghai Jiao Tong University, China Bio
Yanwei Li, Assistant Professor at School of Artificial Intelligence, Shanghai Jiao Tong University. Before that, he was a Senior Research Scientist on multi-modal model at ByteDance Seed, USA. He received Ph.D. degree from the Chinese University of Hong Kong in 2024. His research focuses on Multi-modal Models and Generative AI for downstream tasks with more than 10K citation. Some highlights include Seed 2.0, Seed 1.8, Seed 1.5-VL, LLaVA-OneVision, Mini-Gemini, LLaMA-VID, and LISA. He severs as Area Chair and Reviewer for several top conference and journals.
Title
Semantic Generative Tuning for Unified Multimodal Models
Abstract
Unified multimodal models aim to integrate visual understanding and generation within a single architecture, yet the two capabilities are still optimized with very different objectives: sparse textual supervision for understanding and dense pixel-level targets for generation. In this talk, I will present Semantic Generative Tuning, a post-training approach that introduces semantic visual targets to better align understanding and generation. Through a systematic study of visual supervision at different semantic levels, we find that high-level semantic objectives, particularly segmentation, provide more effective learning signals than low-level reconstruction. SGT consistently improves both visual understanding and generation, suggesting that the choice of visual supervision is critical for building more capable unified multimodal models.
JeongRae (from AISTAT, Top team of MOSEv2 Challenge)
Chung-Ang University, South Korea
Title
SAM3Dual: Training-Free Dual Memory for Long-Term Video Object Segmentation
Abstract
We present SAM3Dual, a training-free approach for long-term video object segmentation. SAM3Dual extends SAM3 with short- and long-term memory branches, confidence-guided memory scaling, and sequence-relative temporal fusion to improve target consistency over time. Without any additional training or adaptation, our method achieved a J&F score of 64.37 and 3rd place in the MOSEv2 track of the 8th LSVOS Challenge at ECCV 2026.
