Scene Graph Compression Using Context-Based Graph Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data compression methods fail to efficiently compress scene graph data, which contains both continuously distributed and discrete data elements, due to their inability to utilize the strong correlations between these elements.
Innovation Solution
A lossless scene graph compression system based on context-based graph convolution, which uses different prediction models with graph convolutional neural networks to estimate prior probabilities for arithmetic encoding and decoding, effectively exploiting contextual information to achieve higher compression ratios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If general-purpose compression methods (gzip, lzma) are used to compress scene graph data, then the compression process is simple, but the compression ratio is low because they ignore contextual correlations between data elements
Solution Approach 1:
The patent changes the parameter representation by transforming scene graph data into probability distributions using graph convolutional networks. Instead of compressing raw data elements directly, the system estimates prior probabilities for each data element based on contextual information from the graph structure, then applies arithmetic coding to these probability estimates, achieving superior compression ratios.
Solution Approach 2:
The patent introduces graph convolutional networks as an intermediary between the scene graph data and the compression algorithm. This intermediary extracts contextual information and generates probability estimates that bridge the gap between the structured graph data and the arithmetic coding process, enabling the compressor to exploit correlations without requiring complex custom logic.
2Adaptability or versatility
If image compression methods (JPEG, MPEG-2) are used for scene graph data, then continuous data can be compressed, but discrete graph structure data cannot be handled effectively
Solution Approach 1:
The patent creates a universal compression framework that handles both continuous and discrete data elements within scene graphs through a unified approach. The graph convolutional network processes the graph structure to generate contextual information that informs probability estimates for all data types, while arithmetic coding provides a unified encoding mechanism that works for both continuous and discrete distributions.
Solution Approach 2:
The patent applies different probability distribution models to different data elements based on their local characteristics. Continuous data elements like object locations use appropriate continuous distributions, while discrete elements like object categories and relationship types use discrete distributions, with each element's compression optimized according to its specific properties and contextual dependencies.
3Adaptability or versatility
If graph compression algorithms (WebGraph, K-squared tree) are used, then discrete graph structure data can be compressed, but continuously distributed data elements cannot be handled
Solution Approach 1:
The patent merges the strengths of graph-based compression with arithmetic coding by combining graph convolutional networks (which excel at processing graph structures) with probability-based arithmetic coding (which excels at compressing both continuous and discrete data). This integration creates a hybrid approach that leverages the graph structure for contextual understanding while using arithmetic coding for efficient compression of all data types.
Data Source
AI summary
A system and method for scene graph lossless compression based on context-based graph convolution comprises a prior encoding module, a first prior decoding module, a data encoding module located at the compressor side, a second prior decoding module, and a data decoding module located at the decompressor side; it extracts prior information from the original scene graph; for different data elements such as object location, object category, interaction relationship, and relation category, it applies different random distribution models with context-based graph convolutional neural networks to estimate the respective prior probabilities for arithmetic encoding and decoding; therefore, it can effectively exploit the contextual information within the scene graph data to perform more accurate probability prediction and thus achieve a better compression ratio.