Large-Scale Texture Identification with Distributed GPU Acceleration for Products Traceability

Texture exists in lots of products, such as wood, beef, and compression tea. These abundant and random texture patterns are significantly different among any two products. Unlike the traditional digital ID tracking, we propose a novel products traceability system using texture identification technology. Similar to face identification, the texture verification and search are both supported. We'll introduce how to build an efficient large-scale texture identification system and how to optimize the fundamental nearest-neighbor algorithm for feature matching with multiple GPU acceleration technologies, such as cuBLAS acceleration, half precision, fined-grained GPU memory planning, and distributed computing. We demonstrated 2.94x speedup and 50% memory saving comparing to the OpenCV kNN CUDA implementation, and the search could be completed in 8.2 seconds from 1 million products with totally 640 Giga feature (128 dim) similarity computation using 14 NVIDIA V100 cards.