Ensemble Detection Models for LiDAR Point Clouds

We address the problem of the ensemble neural networks in the LiDAR pointclouds. Neural networks are sometimes saturated for specific situations meaning they perform worse on some scenarios due to capacity issues. The prediction results in the LiDAR point cloud domain are deteriorating with increasing distance due to the low density of the measurement in the remote areas. We are using a multi-view ensemble, which consists of detection models operating on a separate view and merging the transformation into one data representation. We are using frontview projection, which is the transformation of the canonical coordinates of the LiDAR point cloud to the spherical coordinates. The second view is the projection of scan points to xy plane called Bird’s Eye View (BEV). In the both projections we merge models focusing on specific areas or distance range. We further exploit semi-supervised learning approach called pseudo-labelling in order to generate labels from the ensemble for baseline improvement. All methods are evaluated on semantic segmentation tasks in autonomous driving scenarios and achieve improvement in terms of IoU against the baseline architecture.