Using Visual Saliency to Improve Human Detection with Convolutional Networks

02/21/2018

∙

by   Vandit Gajjar, et al.

∙

0

∙

share

In this paper, we demonstrate an approach based on visual saliency for detection of humans. Using Deep Multi-Layer Network [1], we find the saliency maps of an image having humans, multiply with the input image and fed to Convolutional Neural Network (CNN). For detection purpose, we train DetectNet on prepared two challenging datasets - Penn-Fudan Dataset and TudBrussels Benchmark. After training, the network learns the mid and high-level features of a human body. We show the effectiveness of our approach on both the tasks and report state-of-the-art performance on PennFudan Dataset with the detection accuracy of 91.4 Benchmark.