FeatLoc: Absolute Pose Regressor for Indoor 2D Sparse Features with Simplistic View Synthesizing

This is the PyTorch implementation of our ISPRS 2022 paper. We introduce FeatLoc, a neural network for end-to-end learning of camera localization from indoor 2D sparse features. Authors: Thuan Bui Bach, Tuan Tran Dinh, Joo-Ho Lee

BibTex Citation

If you find this project useful, please cite:

@article{BACH202250,
title = {FeatLoc: Absolute pose regressor for indoor 2D sparse features with simplistic view synthesizing},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
volume = {189},
pages = {50-62},
year = {2022},
doi = {https://doi.org/10.1016/j.isprsjprs.2022.04.021},
}

About

Network Architecture

Results on 7Scenes and 12Scenes

Comparison of average median localization errors

Name	PoseNet17	MapNet	FeatLoc(ours)
7Scenes	0.24m, 8.12°	0.21m, 7.77°	0.14m, 5.89°
12Scenes	0.74m, 6.48°	0.63m, 5.85°	0.38m, 5.04°

Documentation

Setup

• The codes are tested along with
  • Python 3.7,
  • Pytorch 1.5.0,
  • PointNet++ lib,
  • Others python packages including matplotlib, pandas, h5py, tqdm, and numpy.
• To directly install these packages, run

sudo pip install -r requirements.txt

• If you are familiar with conda environments, please run

conda create -f environment.yml
conda activate FeatLoc

• Note that PointNet++ lib needs to install seperately.

Data

1. Install the hierarchical localization toolbox(hloc) into the dataset folder, then change its name to Hierarchical_Localization as bellow.

FeatLoc
├── dataset
│   ├── Generated_Data
│   ├── Hierarchical_Localization
|   ├── gendata.py
|   └── gendata_lib.py
├── model
│   ├── ...
├── ...
└── README.md

2. Generate 3D model

• For 7scenes dataset, please process following this guideline.

• For 12Scenes dataset, please run dsac setup to download the dataset, then use hloc for generating 3D model for each scene. Note that you need to create a 3D model of entire train and test data per scene using Colmap first, then use hloc for only training data.

3. Generate training and testing data.

• Please use the same environment with hierarchical localization toolbox for this part.

cd dataset
python gendata.py --dataset 7scenes --scene chess --augment 1

Running the code

Demo/Inference

• Please run the executable script eval.py for evaluating each scene independently . For example, we can evaluate FeatLoc++ on apt1_living scene as follows:

python eval.py --scene apt1_living --checkpoint results/apt1_living_featloc++au.pth.tar --version 2

Median error in translation = 0.2601 m
Median error in rotation    = 3.8867 degrees

• You can download the prepared testing data and trained models of 12scenes from the Google drive (please move the data folders and model files to dataset/Generated_Data and results folder respectively)

Train

• Please run the executable script train.py to train each scene independently. For example, we can train FeatLoc++ on apt1_living scene as follows:

python train.py --scene apt1_living --n_epochs 200 --version 2 --augment 1

Further enquiry

If you have any problem in running the code, please feel free to contact me: thuan.aislab@gmail.com