Unsupervised Motion Segmentation for Neuromorphic Aerial Surveillance

This is the official repository for Unsupervised Motion Segmentation for Neuromorphic Aerial Surveillance by Sami Arja, Alexandre Marcireau, Saeed Afshar, Bharath Ramesh, Gregory Cohen

Project Page	Paper	Poster

If you use this work in your research, please cite it:

@misc{arja_unsupervised_2024,
	title = {Unsupervised Motion Segmentation for Neuromorphic Aerial Surveillance},
	url = {http://arxiv.org/abs/2405.15209},
	publisher = {arXiv},
	author = {Arja, Sami and Marcireau, Alexandre and Afshar, Saeed and Ramesh, Bharath and Cohen, Gregory},
	month = may,
	year = {2024},
}

Setup

Requirements

• python: 3.8.x, 3.9.x, 3.10.x

Tested environments

• Ubuntu 22.04
• Conda 23.1.0
• Python 3.8.18

Installation

git clone https://github.com/samiarja/ev_deep_motion_segmentation.git
cd ev_deep_motion_segmentation
conda create --name ev_motion_segmentation python=3.9
conda activate ev_motion_segmentation
python3 -m pip install -e .
conda install -c pytorch pytorch torchvision
pip install tqdm
pip install plotly
pip install scikit-image
pip install loris
pip install einops
pip install PyYAML
pip install opencv-python
conda install -c conda-forge pydensecrf

Download dataset

You can download all the dataset from google drive The structure of the folder is as follows:

(root)/Dataset/
        EV-Airborne/
            (sequence_name1).es
            (sequence_name2).es
            (sequence_name3).es
            .....
        EV-IMO/

        EV-IMO2/

        DistSurf/

        HKUST-EMS/

        EED/

Run

Setup the config file

Please see the ./config/config.yaml for an example on how to setup the initial parameters.

Modify the entries to specify the dataset and seq and other parameters.

The seq name can be extracted from the .es file. If the filename is: EED_what_is_background_events.es

Then seq name is what_is_background. It is always between the dataset name (e.g. EED) and events. I will make this easier in future commits.

Execute motion segmentation

python main.py

The output from every layer of the network is saved in subfolders in ./output in this format:

input_frames
RAFT_FlowImages_gap1
RAFT_Flows_gap1
coarse
bs
tt_adapt
rgb
motion_comp
motion_comp_large_delta
config_EV-Airborne_recording_2023-04-26_15-30-21_cut2.yaml
EV-Airborne_recording_2023-04-26_15-30-21_cut2_events_with_motion_inter.h5
motion_segmentation_network_EV-Airborne_recording_2023-04-26_15-30-21_cut2.gif

Description of the content of each subfolder:

• input_frames: The original time surface frames.
• RAFT_FlowImages_gap1: Optical flow images generated by RAFT.
• RAFT_Flows_gap1: The optical flow data in .flo format.
• coarse: The initial output from TokenCut which use the optical flow and the event time surface.
• bs: The output of the bilateral solver which is a binary mask for each frame. If crf: true, then a crf folder will be created.
• tt_adapt: The output from dynamic mask refinement that smooth and correct the mask from the bs or crf.
• rgb: Overlay between the timesurface frame and the tt_adapt.
• motion_comp The output of the motion segmentation.
• motion_comp_large_delta The motion segmentation output but with a higher integration time \delta t.
• config_EV-Airborne_recording_2023-04-26_15-30-21_cut2.yaml The config file will be saved from ./config/config.yaml.
• EV-Airborne_recording_2023-04-26_15-30-21_cut2_events_with_motion_inter.h5: The event with the continuous (i.e. motion) and discrete label (i.e. class id) it has the following structure: 'x','y','p','t','l','cl', 'vx', 'vy'. vx and vy are the continuous motion labels and cl is the discrete label. Both are used to generate the motion segmentation output.
• motion_segmentation_network_EV-Airborne_recording_2023-04-26_15-30-21_cut2.gif: A video gif combining to show input and output of the motion segmentation. Plus there is gif file inside of each subfolder.

A faster implementation is also provided in main_fast_single_object.py, it only works if there for a single moving object.

python main_fast_single_object.py

Acknowledgement

This code is built on top of TokenCut, DINO, RAFT, and event_warping (our previous work). We would like to sincerely thanks those authors for their great works.