SceneCode: Monocular Dense Semantic Reconstruction using Learned Encoded Scene Representations
This addresses the issue of poor semantic representation in monocular dense semantic reconstruction systems, which is incremental as it builds on existing mapping frameworks.
The paper tackles the problem of noisy and incoherent semantic label fusion in 3D mapping by introducing a compact, learned latent space representation using a variational auto-encoder conditioned on colour images, resulting in consistent fused label maps that preserve spatial correlation.
Systems which incrementally create 3D semantic maps from image sequences must store and update representations of both geometry and semantic entities. However, while there has been much work on the correct formulation for geometrical estimation, state-of-the-art systems usually rely on simple semantic representations which store and update independent label estimates for each surface element (depth pixels, surfels, or voxels). Spatial correlation is discarded, and fused label maps are incoherent and noisy. We introduce a new compact and optimisable semantic representation by training a variational auto-encoder that is conditioned on a colour image. Using this learned latent space, we can tackle semantic label fusion by jointly optimising the low-dimenional codes associated with each of a set of overlapping images, producing consistent fused label maps which preserve spatial correlation. We also show how this approach can be used within a monocular keyframe based semantic mapping system where a similar code approach is used for geometry. The probabilistic formulation allows a flexible formulation where we can jointly estimate motion, geometry and semantics in a unified optimisation.