Reproduce “ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weijie Wang, Donny Y. Chen, Zeyu Zhang, Duochao Shi, Akide Liu, Bohan Zhuang
ZPressor, a lightweight and architecture-agnostic module, enables feed-forward 3D Gaussian Splatting models to scale efficiently to dense multi-view inputs by compressing redundant information. This approach improves rendering quality, with PSNR gains up to 4.65 dB, while significantly reducing memory usage by approximately 80% and inference time by 70% in high-density scenarios.
Key insights
- Scalability issues in feed-forward 3DGS with dense inputs stem from information redundancy and representational overload, which can be effectively addressed through information compression.
- The Information Bottleneck principle provides an effective theoretical framework for learning compact representations that retain task-relevant information while discarding redundancy in multi-view data.
- An adaptive selection and fusion strategy, using anchor views and cross-attention, can efficiently manage information density and enable existing models to handle significantly more input views.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.