Accelerating Video Object Detection by Exploiting Prior Object Locations

Abstract

We provide a set of generic modifications to improve the execution efficiency of single-shot object detectors by exploiting prior object locations in video sequences. We propose a crop-based method to accelerate object detection tasks. It dynamically generates crop regions based on prior information and exploits scene sparsity enabling focused use of computational resources. In contrast to prior work, smaller input resolutions for processing crop regions are used to further reduce computational load. The execution efficiency is increased by avoiding multiple executions of the detector in full resolution. Data augmentations are used to successfully train these lower-resolution networks and maintain their accuracy at the baseline level while reducing inference time. Experiments with two public datasets, UA-DETRAC [13] and UAVDT [2], using the SSD-ML [19] object detection architecture with $128\times 128$, $64\times 64$ and $32\times 32$ input resolutions show that we can achieve a maximum speedup by a factor of 1.7 on the UA-DETRAC dataset, and 1.6 on the UAVDT dataset while delivering the same level of accuracy as the base method. An extensive set of experiments demonstrates the speed-accuracy trade-off and shows that our method can achieve accuracy comparable to state-of-the-art methods at lower execution time.

This work is funded by the NWO Perspectief program ZERO.