Learning robust features for indoor scene recognition

Abstract

As digital image acquisition becomes ubiquitous in recent years, the need for indoor scene recognition becomes more pronounced. Existing methods leverage the features of composing objects in a scene and overlook the adverse impacts of the common objects reoccurring in other scenes. This drawback decreases the feature discrimination between scenes (e.g., living room, dining room, and bedroom) due to reoccurring objects (e.g., tables, chairs, and toys). We propose a method of training convolutional networks by punishing or discounting the local object representations’ predictive ability and encouraging the network to learn global scene layout representations. To retain more vital information for the scene feature representation, we introduce an activation function (with unbounded above, bounded below, smooth, and non-monotonic properties) to allow more low-negative values to flow through the network, discarding high negative values. We evaluate the proposed methods on MIT Indoor 67 and Scene 15 datasets. The experiment findings show that the proposed methods capture global scene concepts and improve performance.