OpenSUN3D

7^th Workshop on Open-World 3D Scene Understanding and Representations

September 8 (morning), 2026 — Room: Malmömässan B, in conjunction with ECCV 2026 in Malmö, Sweden.

Introduction

For intelligent agents to thrive in the physical world, they must not only perceive but also comprehend and act within it. This year’s workshop unites leading researchers advancing world models and spatial intelligence, from 3D perception and scene understanding to generative world representations. Through keynotes and a challenge on open 3D interaction perception, we will explore how machines learn to reason about and engage with their environments. The workshop aims to shape the next generation of spatially grounded AI, bridging scientific discovery, practical deployment, and responsible innovation.

Schedule

Workshop Room: Malmömässan B

Poster Presentation Room: Malmömässan Exhibit Hall

09:00 - 09:10	Welcome & Introduction
09:10 - 09:40	Keynote 1 Matthias Nießner (TUM / SpAItial AI)
09:40 - 10:10	Keynote 2 Dima Damen (University of Bristol / Google DeepMind)
10:10 - 11:00	Coffee Break & Poster Session Poster Presentation Room: Malmömässan Exhibit Hall
11:00 - 11:30	Keynote 3 Carl Doersch (Google DeepMind)
11:30 - 12:00	Keynote 4 Peter Kontschieder (Meta)
12:00 - 12:30	Winners Talks
12:30 - 12:40	Concluding Remarks

Keynote Speakers

Matthias Nießner

TUM / SpAItial AI

Dima Damen

University of Bristol / Google DeepMind

Carl Doersch

Google DeepMind

Peter Kontschieder

Research Director at Meta

Poster Session

Poster Presentation Room: Malmömässan Exhibit Hall.
Poster board IDs: 39–58. Any of these may be used for the 10:00 am–11:00 am poster presentation.
Poster boards measure 140 cm wide × 100 cm high (landscape orientation).
Download the ECCV 2026 poster template (.pptx).
Poster printing information will be posted on the website shortly.

Paper Track Closed

Submission Type	Length	In Proceedings	Deadline	Policies
Proceedings Track	14 pages	Yes	August 1, 2026	Anonymous, subject to dual-submission rules.
Non-Archival Track	7 / 14 pages	No	August 1, 2026	Anonymous, may include previously published work.

Submit via OpenReview

All submissions must be anonymous and follow the ECCV 26 author guidelines.

Submission Deadline 1:
Proceedings Track

14-page full papers for inclusion in the ECCV workshop proceedings. Those submissions are subject to the dual submission policy, only original work not under review or published elsewhere is allowed.

Submission deadline: August 1, 2026
Notification to authors: August 10, 2026
Camera-ready submission: August 15, 2026

Submission Deadline 2:
Extended Abstracts & Non-Archival Papers

7-page extended abstracts or 14-page full papers that will not be included in the proceedings. May include previously published work that the authors want to present at the workshop.

Submission deadline: August 1, 2026
Notification to authors: August 10, 2026
May include previously published work.

★ NEW

Huawei SpatialMem Track

A dedicated sub-track on memory-centric, online, dynamic/4D or world-model-coupled spatial intelligence. Includes both archival and non-archival submissions. Authors may submit to the main workshop track and also flag their paper as SpatialMem for consideration in this sub-track. Submit as proceedings or non-archival exactly as above and simply flag your paper as SpatialMem at submission.

🏆 Best Paper Award & a dedicated Best SpatialMem Paper Award funded by Huawei.

Topics of interest ▾

Dynamic & Hierarchical Spatial Representations — persistent open-vocabulary 3D scene graphs (metric, semantic, topological, 4D) and multi-view-consistent, metric-grounded VLM reasoning.
Test-Time Memory Evolution & Updating — updating, editing and bounding a 3D memory at inference across revisits; consolidation, selective forgetting, and memory for video world models.
Neuro-Symbolic Spatial Reasoning — verifiable programs, tools and graph-query interfaces over 3D scene graphs for checkable inference.
Systems & Efficiency — on-device, power/memory-budgeted mapping + reasoning; constant-memory streaming maps and deployment-overhead benchmarks.
Multi-Session Persistent Spatial Memory — representations and benchmarks for memory persisting across sessions (hours to days+), tracking identity, relations, uncertainty and provenance.

Challenge Track Closed

We host a 3-track challenge based on the SceneFun3D benchmark and the Articulate3D dataset. The challenge focuses on fine-grained functionality, affordance, and interaction understanding in 3D indoor environments.