A spacetime model for one-shot active contour extraction scheme for human detection in image sequences

Abstract

We present a simple yet effective 3D convolutional neural network that learns a novel spacetime representation for human contour detection in a sequence of images. Our approach, in one shot, detects the contour of humans while generating high-quality results compared to the traditional binary mask representations. Our time-consistent convolutional neural network takes a sequence of images as its input and generates an implicit level set surface, in which the object boundaries correspond to the zero-level set. Furthermore, we introduce an appropriate way to combine space and time in an interwoven coordinate system tailored to spatiotemporal datasets. We showcase the feasibility of our approach by training the network on a semi-synthetic dataset. We discuss various configurations of our approach, all of which is shown to outperform the typical binary mask representation. We believe that this new approach could potentially improve the performance of all architectures compared to their alternative ones. The code will be made available on https://github.com/orgs/OSUPCVLab/HumanContourDetection.