Context-Aware Object Detection With Convolutional Neural Networks

Yizhou Yan, Lei Cao, Samuel Madden, Elke Rundensteiner

TL;DR: A deep neural network that leverages conditional random field to enforce context semantics constrains in object detection

Abstract: Although the state-of-the-art object detection methods are successful in detecting and classifying objects by leveraging deep convolutional neural networks (CNNs), these methods overlook the semantic context which implies the probabilities that different classes of objects occur jointly. In this work, we propose a context-aware CNN (or conCNN for short) that for the first time effectively enforces the semantics context constraints in the CNN-based object detector by leveraging the popular conditional random field (CRF) model in CNN. In particular, conCNN features a context-aware module that naturally models the mean-field inference method for CRF using a stack of common CNN operations. It can be seamlessly plugged into any existing region-based object detection paradigm. Our experiments using COCO datasets showcase that conCNN improves the average precision (AP) of object detection by 2 percentage points, while only introducing negligible extra training overheads.

Keywords: Object Detection, CNN, Context, CRF