Incremental instance-level 3D segmentation performed by fusing geometric and deep-learning cues

This repository contains code from my master project at University College London. It aims at incrementally achieving an object-aware 3D segmentation by fusing geometrical and deep-learning frame-wise information. It's build on InfiniTam (https://github.com/victorprad/InfiniTAM).

Getting Started

Prerequisites

Everything InfiniTam needs

OpenCV

Build Process

  $ mkdir build
  $ cd build
  $ mkdir normals
  $ cmake ../InfiniTAM/ 
  $ make

Run the code

Run with other images

To test with images, the image folder needs to be in build and use the same format as InfiniTam (calib file and a 'frames' folder)

To test without semantic segmentation use :

./InfiniTAM folderName/calib.txt folderName/Frames/%04i.ppm folderName/Frames/%04i.pgm

and with semantic segmentation :

./InfiniTAM folderName/calib.txt folderName/Frames/%04i.ppm folderName/Frames/%04i.pgm folderName/Frames/sem_seg%04i.pgm

Saving results

The code to save result is in UIEngine.cpp, result folders need to be added to the sequence folder. For example :

build
  sequence
    frames
    calib.txt
    out

Thanks !