SmoothMotionVectors Stores Free-Viewpoint Video in Less Space

SmoothMotionVectors makes deformation signals easier for standard video codecs to compress, reducing free-view video storage by about 20 percent while preserving quality.

Paper 《SmoothMotionVectors: Optimizing Your Content for Video Codecs in Free View Video Compression》

Authors Mingyang Song · Yang Zhang · Siyu Tang · Tunç Ozan Aydin

Research Disney Research Studios · ETH Zurich

Presented at SIGGRAPH 2026

Release 2026-07-16

Video 9 minutes 41 seconds

Adapting 3D Gaussian Splatting Scenes to Video Codecs

《SmoothMotionVectors》 is a study analyzing pipeline choices that affect quality and file size when dynamic scenes reconstructed with 3D Gaussian Splatting are stored as free-viewpoint video. It compares temporal chunking, deformation-based reconstruction, and quantization-aware training one by one to identify conditions that improve real-world compression efficiency.

Using a short GOP reduces the accumulation of errors between frames and improves reconstruction quality, but increases storage requirements because more keyframes are needed. The research team measures this trade-off and proposes smoothing the vectors that represent scene motion across both space and time.

Smooth motion vectors create deformation signals that existing video codecs can compress more easily. Instead of designing a dedicated codec, this approach refines the content representation so that widely used compression tools can efficiently handle the temporal changes in 3D scenes as well.

In experiments, storage space was reduced by approximately 20% while maintaining or improving image quality compared with existing state-of-the-art methods. The practical strength of the study is that it considers not only the quality of free-viewpoint video but also the real costs of transmission and storage.

🏷 Other posts for the same work →

Maru