The paper proposes Pointsoup, an efficient learning-based geometry codec for compressing large-scale point cloud scenes. Key highlights:
Pointsoup leverages a point model-based strategy to characterize local surfaces. It embeds skin features from local windows via an attention-based encoder, and introduces dilated windows as cross-scale priors to infer the distribution of quantized features in parallel.
During decoding, Pointsoup employs a fast feature refinement block followed by an efficient folding-based point generator to reconstruct the local surface with fast speed. This enables extremely low decoding latency, up to 90-160x faster than the G-PCCv23 Trisoup decoder on a single RTX 2080Ti GPU.
Pointsoup achieves state-of-the-art compression performance on multiple benchmarks, providing 60-64% bitrate reduction over the G-PCCv23 anchor. It also offers flexible bitrate control with a lightweight neural model (2.9MB), which is beneficial for practical applications.
Experiments show that Pointsoup can effectively handle large-scale point cloud scenes, demonstrating strong generalization capability from a small-scale training dataset.
To Another Language
from source content
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Kang You,Kai... lúc arxiv.org 04-23-2024
https://arxiv.org/pdf/2404.13550.pdfYêu cầu sâu hơn