Scene Signatures: Localised and Point-less Features for Localisation

Abstract

Abstract-This paper is about localising across extreme lighting and weather conditions. We depart from the traditional point-feature-based approach since matching under dramatic appearance changes is a brittle and hard. Point-feature detectors are rigid procedures which pass over an image examining small, low-level structure such as corners or blobs. They apply the same criteria to all images of all places. This paper takes a contrary view and asks what is possible if instead we learn a bespoke detector for every place. Our localisation task then turns into curating a large bank of spatially indexed detectors and we show that this yields vastly superior performance in terms of robustness in exchange for a reduced but tolerable metric precision. We present an unsupervised system that produces broad-region detectors for distinctive visual elements, called scene signatures, which can be associated across almost all appearance changes. We show, using 21 km of data collected over a period of 3 months, that our system is capable of producing metric estimates from night-to-day or summer-to-winter conditions.

Image features for visual teach-and-repeat navigation in changing environments

“…For example, Churchill and Newman [9] clustered different observations of the same place to form "experiences" that characterize the place appearance in particular conditions. McManus et al [38] used dead reckoning to predict which place the vehicle is close to, loaded a bank of Support Vector Machine classifiers associated with that place and used these to obtain a metric pose estimate. Krajnik et al [39] proposed to maintain maps gathered over an entire year and select the most relevant map based on its mutual information with the current observation.…”

Absolute geo-localization thanks to Hidden Markov Model and exemplar-based metric learning

“…In [11], the problem is cast as a classification task, a classifier for each image in the database is trained using per-exemplar SVM approach. In [21] the authors propose to learn a bank of detectors for each place, to identify specific scene structures instead of local features.…”

Only look once, mining distinctive landmarks from ConvNet for visual place recognition

“…Instead, representations that break down an image into smaller regions, such as [15] can be more robust against scale and viewpoint variations. The approaches in [7,8] combine an external landmark detector with CNN-based features to match regions over extreme viewpointand condition-variations.…”