Image-based localization using LSTMs for structured feature correlation
This work addresses image-based localization for robotics and AR/VR applications, offering a novel hybrid approach that is incremental in combining CNNs with LSTMs for structured feature correlation.
The authors tackled camera pose regression in indoor and outdoor scenes by proposing a CNN+LSTM architecture that improves localization performance, particularly in challenging conditions like textureless surfaces, outperforming existing deep methods and showing robustness against motion blur and illumination changes.
In this work we propose a new CNN+LSTM architecture for camera pose regression for indoor and outdoor scenes. CNNs allow us to learn suitable feature representations for localization that are robust against motion blur and illumination changes. We make use of LSTM units on the CNN output, which play the role of a structured dimensionality reduction on the feature vector, leading to drastic improvements in localization performance. We provide extensive quantitative comparison of CNN-based and SIFT-based localization methods, showing the weaknesses and strengths of each. Furthermore, we present a new large-scale indoor dataset with accurate ground truth from a laser scanner. Experimental results on both indoor and outdoor public datasets show our method outperforms existing deep architectures, and can localize images in hard conditions, e.g., in the presence of mostly textureless surfaces, where classic SIFT-based methods fail.