Skip to content

gokulbnr/fast-slow-biased-event-vpr

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

98 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Enhancing Visual Place Recognition via Fast and Slow Adaptive Biasing in Event Cameras

Project Video Website Dataset Preprint

Welcome to the official repository for the paper Enhancing Visual Place Recognition via Fast and Slow Adaptive Biasing in Event Cameras, to be presented at the 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024). This work introduces feedback control algorithms that dynamically change bias parameters for event-cameras to stabilize event-rate in an online manner. The work reports improvements in visual place recognition performances across variations in environment brightness conditions, validated through comprehensive real-time evaluations using a new QCR-Fast-and-Slow-Event-Dataset.

First GIF Image Second GIF
With Default Biases Our Bias Control Approach With Proposed Approach

Getting Started with Mamba

Borrowing the story from Robostack: To get started with conda (or mamba) as package managers, you need to have a base conda installation. Please do not use the Anaconda installer, but rather start with miniforge that is much more "minimal" installer. This installer will create a "base" environment that contains the package managers conda and mamba. After this installation is done, you can move on to the next steps.

When you already have a conda installation you can install mamba with:

conda install mamba -c conda-forge

Code and Environment Setup

git clone [email protected]:gokulbnr/fast-slow-biased-event-vpr.git
cd fast-slow-biased-event-vpr
mamba env create -f environment.yaml
mamba activate evpr
pip install git+ssh://[email protected]/gokulbnr/tonic.git@develop

Fast and Slow Biasing of Event Cameras (During Dataset Collection)

We make use of a ROS1 node fast_and_slow_controller to update Event-Camera bias parameters in a online manner. This node has to run with jAERv1.9.5 with unicast datagram (UDP) output enabled.

Prerequisite: jAER

To sort out your dependencies for jAER, please use its user guide. The proposed approach has been rigorously tested on devices running Ubuntu 20 and Ubuntu 22, both utilizing x86_64 architecture. The instructions to setup jAER up is as follows:

git clone [email protected]:SensorsINI/jaer.git
cd jaer
git checkout 1.9.5
time ant jar

Setting up Fast and Slow Bias Controller ROS Node

cd fast-slow-biased-event-vpr
mkdir -p ~/catkin_ws/src/
mv fast_and_slow_controller_ros ~/catkin_ws/src/
cd ~/catkin_ws/
mamba activate evpr
catkin build

Dataset Collection

The event streams were recorded using the jAERv1.9.5 library for a DAVIS346Red device. jAER was run with unicast datagram (UDP) output enabled to ensure UDP communication between the fast_and_slow_controller ROS node and jAER is enabled. To run the ROS node:

cd ~/catkin_ws/
source devel/setup.bash
rosrun fast_and_slow_controller_ros fast_and_slow_controller

Data Processing

Link to the released dataset: https://huggingface.co/datasets/gokulbnr/QCR-Fast-Slow-Event-Dataset

The dataset contains raw DVS data in AEDAT2.0 format and ground truth poses of the moving camera (pose of robot on which the DAVIS346Red camera is mounted) as tf2 transforms in rosbag files. To process raw data from traverses into geotagged image sequences, please use scripts/process_data.sh.

mamba activate evpr
cd fast-slow-biased-event-vpr/event_vpr
bash scripts/process_data.sh <experiment_name> <number_of_iterations> <path_to_experiment_home> <save_path_for_processed_data>

Visual Place Recognition (Testing Data on Downstream Task)

We tested with Sum of Absolute Differences (SAD) to perform similarity computations between reference and query sets.

mamba activate evpr
cd fast-slow-biased-event-vpr/event_vpr
bash scripts/run_vpr.sh <experiment_name> <save_path_for_results> <path_to_processed_data_root_directory> <brightness_condition>

The experiment_names follow the same nomenclature as that in the main tables of the paper manuscript. PxBw, PxTh, RfPr were available within jAER in the list of Filters (bottom left tab on the GUI). They are available under the DVSBiasController Filter in jAER. default_params corresponds to jAER's constant bias settings when photoreceptor bandwidth, event threshold, and max pixel firing rate parameters are all at 0. These parameters can be found under "User Friendly Settings" within "HW Configuration" (bottom left tab on GUI). Fast_Slow corresponds to results using our Fast and Slow Bias Controller. Here's a list of permitted values for arguments <data_root>, <experiment_name>, brightness_condition, and iteration.

Data Root Directory Names Experiment Names Brightness Conditions Number of iterations
main_experiments Fast_Slow low, medium, high 5
default_params low, medium, high 5
PxBw low, medium, high 5
PxTh low, medium, high 5
RfPr low, medium, high 5
ablation_study_components constant low, medium, high 3
fast low, medium, high 3
slow low, medium, high 3
ablation_study_slow_changes_freq n2 low, medium, high 3
n7 low, medium, high 3
n10 low, medium, high 3

Cite us at

@inproceedings{nair2024enhancing,
  title={Enhancing Visual Place Recognition via Fast and Slow Adaptive Biasing in Event Cameras},
  author={Nair, Gokul B and Milford, Michael and Fischer, Tobias},
  booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2024}
}

License

This project is licensed under the MIT License. See the LICENSE file for details.