Improved Bilinear Pooling for Real-Time Pose Event Camera Relocalisation

Traditional methods for estimating camera pose have been replaced by more advanced camera relocalization methods that utilize both CNNs and LSTMs in the field of simultaneous localization and mapping. However, the reliance on LSTM layers in these methods can lead to overfitting and slow convergence. In this paper, a novel approach for estimating the six degree of freedom (6DOF) pose of an event camera using deep learning is presented. Our method begins by preprocessing the events captured by the event camera to generate a set of images. These images are then passed through two CNNs to extract relevant features. These features are multiplied using an outer product and aggregated across different regions of the image after adding L2 normalization to normalize the combining vector. The final step of the model is a regression layer that predicts the position and orientation of the event camera. The effectiveness of this approach has been tested on various datasets, and the results demonstrate its superiority compared to existing state-of-the-art methods.