Predicting Matchability

Will It Last? Learning Stable Features for Long-Term Visual Localization

“…Similarly to [9], the approach presented by Kim et al [16] aims at detecting features that will have high matching score for obtaining good geo-localization. Turcot and Lowe [37] propose an approach that requires several different views of a scenery to be present in order to asses the quality of the features therein.…”

“…Including the depth information brings a benefit over using the raw image only. the-art [9]). A top-down view of the localization results is presented in Figure 10.…”

“…Predicting Matchability score: We also used a method that was proposed in [9]. It is expected to select wellmatchable and uniquely-matchable features.…”

“…While that work is useful for map maintenance, it does not aim at providing an a priori feature quality criterion. In the work by Hartmann et al [9], a random forest is trained to predict the matchability of hand-crafted SIFT descriptors. That is, the classifier predicts feature distinctiveness for better matching between consecutive frames but does not detect whether a feature (with a potentially high matchability score) might belong to a dynamic object or have an unstable appearance over time.…”

“…Some approaches use the past data to find features which are reliably redetected [31], but ignore their actual semantic meaning. Other methods attempt to classify hand-crafted descriptors [9] to guarantee a high probability of matching. But even if a feature is well-recognizable and a good matching seems likely, it might be unreliable for life-long localization, e.g., because it belongs to an object that is moved frequently or may even entirely disappear (such as a pen or a mug on a desk, see Figure 1).…”