TopoGPU streams image topology calculations through GPUs
In plain words
Persistent homology is a way to track how image features, such as connected regions and loops, appear and disappear as the image is viewed at different scales. TopoGPU is a method that processes image chunks on a GPU, then finishes a final reduction on the CPU.
Across the image datasets tested, its full computation averaged 53.24 times faster than CRipser, a CPU method, on the authors’ RTX 4080 system. The result applies to those tests and that setup.
The research
Can persistent homology, which tracks how image features such as loops appear and disappear across scales, be used in GPU-based image workflows without the cost of moving intermediate data to the CPU? That transfer and synchronization overhead can make the computation a training bottleneck.
TopoGPU streams image chunks through GPU-based topology calculations, then reduces the accumulated boundary matrix on the CPU to compute persistence diagrams.
The authors test whether image-based persistent homology can be computed faster without changing the topological information it returns. TopoGPU processes image chunks on the GPU, accumulates their boundary relations, and leaves the final matrix reduction to the CPU.
Across the evaluated image datasets, full runs averaged 53.24 times faster than CRipser, measured from data loading through matrix reduction on an RTX 4080 system.
A proof that the streamed construction preserves persistent homology up to zero-persistence pairs supports interpreting this as a faster computation of the intended result.
The speed results cover the evaluated image datasets and hardware, not every input or system. For some very large volumes, CPU matrix reduction fails when the full matrix exceeds available RAM; on very small images, setup and data-transfer overhead can let a CPU method finish sooner.
Analysis
Deep
- Authors
- Fan Wang, Hubert Wagner, Rezaul Chowdhury, Chao Chen
- License
- arXiv non-exclusive