engleski

Robust Traffic Scene Recognition with a Limited Descriptor Length

In this paper we describe a novel image descriptor designed for classification of traffic scene images in fleet management systems. The descriptor is computationally simple and very compact (as short as 48 bytes). It is derived from variations of two well known image descriptors: GIST and spatial Fisher vectors, thus encoding both global and local image features. Both GIST (being a global scene descriptor) and spatial Fisher vectors (that relies on local image features) are tuned to produce very short outputs (64 components), which are then concatenated. The output is further compressed by a lossy encoding scheme, without sacrificing classification performance. The encoding scheme uses as little as 3 bits to encode each vector component. The descriptor is evaluated on the publicly available FM2 dataset of traffic scene images. We demonstrate very good classification performance matching that of full-sized general purpose image descriptors.

scene classification; short image descriptors; GIST; spatial Fisher vectors

nije evidentirano