This repository acquires continuous frame sequences through cameras mounted on drones and uses an end-to-end feedforward deep learning architecture to directly regress three-dimensional geometric representations from two-dimensional images.

A unified, feed-forward model integrating the MapAnything 3D reconstruction backbone with Mask2Former for multi-view consistent panoptic segmentation. This approach provides feature-rich and metric accurate semantic maps for indoor robot navigation while eliminating the need for posed inputs and costly test-time optimization.