Reproduce “Active Neural Mapping” on OpenResearch

About this paper

Zike Yan, Haoxiang Yang, Hongbin Zha

Active Neural Mapping from Peking University and PKU-SenseTime Machine Vision Joint Lab introduces an online active mapping system that leverages continual learning of implicit neural representations (INRs) to efficiently explore previously unseen 3D environments. It quantifies map uncertainty through 'neural variability' by perturbing network weights and consistently achieves higher scene completeness compared to traditional active mapping techniques.

Key insights

• Uncertainty in implicit neural representations (INRs) can be quantified by 'neural variability,' which measures the robustness of predictions to small perturbations in the network's weights. High variability indicates high uncertainty.
• Active mapping can be framed as an optimization problem within the weight space of a continually-learning neural network, where the network's internal dynamics directly inform exploration strategy.
• The process of continual learning for an INR, balancing generalization to new areas and avoiding catastrophic forgetting of observed regions, inherently provides signals for identifying areas requiring further data.

Summary by alphaXiv.