IEEE ICME 2026 Grand Challenge
DyGS-Bench: Dynamic Gaussian Splatting Compression & Reconstruction
Important Dates (tentative)
2026.01.21
Website information due to ICME'26 (Word document)
2026.02.10
Starter kit + baseline release (scripts, format checks)
2026.02.15
Development phase opens (online submissions)
2026.04.15
Development phase submission deadline
2026.05.05
Testing phase opens
2026.05.10
Final submission deadline
2026.05.15
Results announcement (recommended by ICME'26 GC chairs)
2026.05.30
ICME'26 Grand Challenge author registration deadline
Challenge Overview
Dynamic 3D Gaussian Splatting (DyGS) enables real-time, high-fidelity rendering of dynamic scenes and is increasingly used in AR/VR/XR and volumetric video. However, naively storing a time-varying set of Gaussian primitives leads to prohibitive storage and bandwidth costs. While many compression approaches have been proposed, progress is slowed by the lack of a standardized benchmark that evaluates
rate
,
rendering quality
, and
online decoding cost
under a unified protocol.
DyGS-Bench
focuses on a realistic deployment setting where only the reconstructed
native Gaussian sequence
is available (no original multi-view training images/videos). The goal of this Grand Challenge is to advance practical codecs that compress dynamic Gaussian assets into a compact bitstream and reconstruct renderable Gaussians efficiently.
Task
Given a native dynamic Gaussian sequence {G_t} and a fixed camera set C, design an
encoder
(offline) and
decoder
(online, measured) that:
Produces a compact
bitstream
from {G_t} (offline encoding is not time-constrained).
Reconstructs renderable Gaussians Ghat_t from the bitstream (online decoding time and memory are measured).
Achieves high rendering quality when rendering Ghat_t under C compared with renderings of {G_t}.
Evaluation (high-level)
Quality:
PSNR / SSIM / LPIPS on rendered images under the fixed camera set.
Rate:
total bitstream size (and derived metrics such as MB/frame or bpp).
Efficiency:
online decoding FPS and peak GPU memory (measured on decoded Gaussians).
Please refer to the Data and Evaluation pages for details on dataset access, formats, and submission rules.
Contact
Forum:
please use the competition forum for Q&A (recommended).
Email:
see the Organizers page for contact emails.