GANmouflage: 3D Object Nondetection with Texture Fields

Rui Guo1 Jasmine Collins2 Oscar de Lima1 Andrew Owens1

1University of Michigan 2UC Berkeley

This repository includes codes for the paper: GANmouflage: 3D Object Nondetection with Texture Fields. arXiv:

Environment Setup

We provide instructions for creating a conda environment for training and generating camouflaged textures.

conda create -n camo_env -y python=3.7
conda activate camo_env
sh ./env.sh

Dataset

  • Scene image data can be downloaded from link. [Owens et al., 2014] Download the data and unzip data into the folder outside the code repository. Make sure scene data is in ../camo-data/ Then run

    python get_num_views.py
    
  • Animal shapes can be downloaded from link. Animal shapes are collected from SMAL. We normalize the size of animals and flipped y-axis to accomodate to our axis definition. Download the data and unzip data into the folder outside the code repository. Make sure animal shape data is in ../fake_animals_v4/

Or directly run

sh ./prepare_data.sh

Training

A sample training command is included in train_ddp.sh.

Scene name can be specified through --scene SCENE_NAME.

If you want to run the method on animal shapes use --animals

Generating Textures

A sample generating command is included in generate.sh.