Fast Integral Histogram Computations on GPU for Real-Time Video Analytics

TL;DR

This paper presents optimized GPU algorithms for fast integral histogram computation, enabling real-time video analytics with significant speedups over CPU implementations, including multi-GPU scaling.

Contribution

It introduces efficient GPU kernel strategies, tiling, and multi-GPU mapping for integral histograms, achieving high frame rates and large data processing capabilities.

Findings

Achieved 300 frames/sec for 640x480 images with 32 bins on a single GPU.

Realized 0.73 Hz processing of 32 GB data using multiple GPUs.

Speedup factors of up to 153x over single-threaded CPU implementations.

Abstract

In many Multimedia content analytics frameworks feature likelihood maps represented as histograms play a critical role in the overall algorithm. Integral histograms provide an efficient computational framework for extracting multi-scale histogram-based regional descriptors in constant time which are considered as the principle building blocks of many video content analytics frameworks. We evaluate four different mappings of the integral histogram computation onto Graphics Processing Units (GPUs) using different kernel optimization strategies. Our kernels perform cumulative sums on row and column histograms in a cross-weave or wavefront scan order, use different data organization and scheduling methods that is shown to critically affect utilization of GPU resources (cores and shared memory). Tiling the 3-D array into smaller regular data blocks significantly speeds up the efficiency of the computation compared to a strip-based organization. The tiled integral histogram using a diagonal wavefront scan has the best performance of about 300.4 frames/sec for 640 x 480 images and 32 bins with a speedup factor of about 120 using GTX Titan X graphics card compared to a single threaded sequential CPU implementation. Double-buffering has been exploited to overlap computation and communication across sequence of images. Mapping integral histogram bins computations onto multiple GPUs enables us to process 32 giga bytes integral histogram data (of 64MB Image and 128 bins) with a frame rate of 0.73 Hz and speedup factor of 153X over single-threaded CPU implementation and the speedup of 45X over 16-threaded CPU implementation.