Skip to content

silviutroscot/CodeSLAM

Repository files navigation

CodeSLAM

PyTorch implementation of CodeSLAM - Learning a Compact, Optimisable Representation for Dense Visual SLAM.

Summary

Problems it tries to tackle/solve

  • Representation of geometry in real 3D perception systems.
    • Dense representations, possibly augmented with semantic labels are high dimensional and unsuitable for probabilistic inference.
    • Sparse representations, which avoid these problems but capture only partial scene information.

The new approach/solution

  • New compact but dense representation of scene geometry, conditioned on the intensity data from a single image and generated from a code consisting of a small number of parameters.
  • Each keyframe can produce a depth map, but the code can be optimised jointly with pose variables and with the codes of overlapping keyframes, for global consistency.

Introduction

  • As the uncertainty propagation quickly becomes intractable for large degrees of freedom, the approaches on SLAM are split into 2 categories:
    • sparse SLAM, representing geometry by a sparse set of features
    • dense SLAM, that attempts to retrieve a more complete description of the environment.
  • The geometry of natural scenes exhibits a high degree of order, so we may not need a large number of params to represent it.
  • Besides that, a scene could be decomposed into a set of semantic objects (e.g a chair) together with some internal params (e.g. size of chair, no of legs) and a pose. Other more general scene elements, which exhibit simple regularity, can be recognised and parametrised within SLAM systems.
  • A straightforward AE might oversimplify the reconstruction of natural scenes, the novelty is to condition the training on intensity images.
  • A scene map consists of a set of selected and estimated historical camera poses together with the corresponding captured images and supplementary local information such as depth estimates. The intensity images are usually required for additional tasks.
  • Depth map estimate becomes a function of corresponding intensity image and an unknown compact representation (referred to as code).
  • We can think of the image providing local details and the code supplying more global shape params and can be seen as a step towards enabling optimisation in general semantic space.
  • The 2 key contributions of this paper are:
    • The derivation of a compact and optimisable representation of dense geometry by conditioning a depth autoencoder on intensity images.
    • The implementation of the first real-time targeted monocular system that achieves such a tight joint optimisation of motion and dense geometry.

Usage

  • generate the python module for the protobuf: protoc --python_out=./ scenenet.proto

Results

Requirements

  • Python 3.4+
  • PyTorch 1.0+
  • Torchvision 0.4.0+