An end-to-end deep-learning method for dense voxel 3D object reconstruction from a single monocular neuromorphic (event) camera that removes the reliance on estimated physical priors and the complex multi-step pipelines used by prior event-based approaches. A novel event representation enhances edge features so a convolutional encoder-decoder feature-enhancement model can learn reconstruction directly, and an Optimal Binarization Threshold Selection Principle is proposed as a benchmark and design guideline, using the best reconstruction under threshold optimization as reference. The approach improves reconstruction accuracy by 54.6% over the baseline.
Neuromorphic (event) cameras are asynchronous brightness-change sensors that capture extremely fast motion without motion blur, making them attractive for 3D reconstruction in extreme environments. P…