Learning Continuous Depth Representation via Geometric Spatial Aggregator