High-Freedom Volume Data Classification Based on Differentiable Volume Splatting

Transfer functions play a crucial role in classifying volume data. However, traditional transfer function design methods require extensive user interactions and often yield inconsistent results. This paper presents a novel approach to volume data classification based on differentiable volume Splatting technology. First, a hierarchical dimensionality reduction method utilizing UMAP and MipMap-like techniques is employed to achieve low-dimensional manifold embeddings of voxels while preserving their spatial topological characteristics, providing users with a dimensionality reduction view for volume data exploration; then, a differentiable volume Splatting renderer is adopted in the rendering view, using differentiable footprint functions to inversely derive voxel contributions to user-defined regions of interest based on rendering outcomes; finally, sensitivity analysis, local transfer function reconstruction, and a series of intuitive interactive tools are integrated to establish visual associations between the dimensionality reduction view and rendering view, forming a “Dual View” system with coactive interactions between dual domains. Experimental results demonstrate that the proposed method achieves dimensionality reduction within 3 s to 5 s while maintaining a rendering frame rate above 60 frames/s, and effectively enhances both the flexibility and reliability of volume data classification compared to traditional methods.