Motion Illusions

Support library for 2020 Telluride Neuromorphic Engineering Workshop Challenge "Insights into the Early Motion Pathway (VISION)"

A description of some of the illusions of interest is available here.

Repository Contents

This library implements helper functions for generating data representative of optical illusions humans experience.

The master branch containers a Python package called motion_illusions. The package contains helper functions for simulating the illusory flow and bias estimation illusions.

The stepping_feet_illusion_matlab branch contains Matlab code for generating events from the stepping feet illusion.

To checkout stepping feet after cloning this repo run:

git fetch origin steppingfeet_illusion_matlab
git checkout steppingfeet_illusion_matlab

The variational-leviant branch contains Matlab code for generating illusory patterns inspired by the Leviant illusion.

To checkout stepping feet after cloning this repo run:

git fetch origin variational-leviant
git checkout variational-leviant

Overview of motion_illusions python package

The library implements image warping, visualization utilities, and wraps the UnFlow optical flow estimator.

The image warping is meant to simulate the small image pertubations due to saccades.

UnFlow is an unsupervised deep optical flow estimator. Currently the library supports using it to evaluate illusory patterns using pre-generated data. The intent is to extend the network to accept multiple frames as would be necessary for enforcing a causality requirement.

Examples

• ./examples
  • rotation_warp_image.py Demonstrate continuously warp an image by a rotation and plotting utilities
  • translation_warp_image.py Demonstrate continuously warp an image by a translation and plotting utilities

Library files

• ./motion_illusions
  • rotation_translation_image_warp.py Utilities for warping an image based on rotation and translation
• ./motion_illusions/utils
  • flow_plot.py Convert optical flow into an image
  • image_tile.py Easily compose multiple images into a single image for display, useful for debugging
  • rate_limit.py Rate limit a process by wall clock time (useful for visualization)
  • signal_plot.py Produce an image plotting multiple real time signals
  • time_iterator.py Iterate over time using a wall clock or fixed timestep

Dependencies

• Python >= 3.5
• CUDA enabled GPU
• See requirements.txt for python dependencies

Setup

Since training neural networks is the goal it assumed a GPU is available.

Clone the repository, initiate submodules.

git clone https://github.com/prgumd/motion_illusions.git
cd motion_illusions
git submodule update --init

Download FlowNet2-SD pretrained model weights and put in motion_illusions/flownet2_pytorch/checkpoints. You will need to create the folder.

Docker

A Dockerfile is provided to build an image with all packages installed.

Some of the code assumes it can display GUIs using an X server. A docker run command is provided for Ubuntu host systems that configures the container for X forwarding.

Make sure docker is installed with access to the GPU.

• Docker
• nvidia-docker

Run the following to build the docker image.

cd motion_illusions
docker build --tag motion_illusions:3.0 .

To launch the container with X forwarding, GPUs available, and the latest version of this package:

cd motion_illusions
docker run  -u $(id -u):$(id -g) -e DISPLAY -v="/tmp/.X11-unix:/tmp/.X11-unix:rw" --ipc host --gpus all --rm -it -v $(pwd):/workdir motion_illusions:3.0 bash

The container will complain about having no user due to overriding the uid and gid for X forwarding. This is ok for our purposes, it is the simplest way to handle X forwarding without a security compromise. More details available here.

On first run the FlowNet2 CUDA modules need to built and installed manually. For some reason this is not working from the Dockerfile yet

cd motion_illusions/flownet2_pytorch
bash install.sh

Virtualenv

Make sure the following is installed on the host.

• CUDA 9.2
• Python >= 3.5
• virtualenv ('pip3 install virtualenv')

Run the following steps to create a virtual environment, activate it, and install packages.

cd motion_illusions
virtualenv venv
source venv/bin/activate
pip install -r requirements.txtc
pip install -e .

Running Flownet2-SD

Inference

FlowNet2-SD - Flying Chairs SD

Assumes Flying Chairs Small Displacement is extracted to motion_illusions/flownet2_pytorch/ChairsSDHom

For some reason EPE is very high with pre-trained weights, something may be wrong with the implementation, more investigation is needed. Full FlowNet2 model has low error as expected.

python3 main.py --inference --model FlowNet2SD --save_flow --inference_dataset ChairsSDHomTrain --inference_dataset_root ChairsSDHom/data --resume checkpoints/FlowNet2-SD_checkpoint.pth.tar

FlowNet2 - Sintel Clean

Assumes Sintel full is downloaded to motion_illusions/flownet2_pytorch/mpi-sintel

python3 main.py --inference --model FlowNet2 --save_flow --inference_dataset MpiSintelClean --inference_dataset_root mpi-sintel/training --resume checkpoints/FlowNet2_checkpoint.pth.tar

Training

FlowNet2-SD - ChairsSDHom - MultiScale

python3 main.py --batch_size 8 --model FlowNet2SD --optimizer=Adam --loss=MultiScale --loss_norm=L1 --loss_numScales=5 --loss_startScale=4 --optimizer_lr=1e-4 --training_dataset ChairsSDHomTrain --training_dataset_root ChairsSDHom/data  --validation_dataset ChairsSDHomTest --validation_dataset_root ChairsSDHom/data

FlowNet2 - Sintel - L1

Assumes Sintel full is downloaded to motion_illusions/flownet2_pytorch/mpi-sintel

python3 main.py --batch_size 8 --model FlowNet2 --loss=L1Loss --optimizer=Adam --optimizer_lr=1e-4 --training_dataset MpiSintelFinal --training_dataset_root mpi-sintel/training   --validation_dataset MpiSintelClean --validation_dataset_root mpi-sintel/training

FlowNet2C - Sintel - MultiScale

python3 main.py --batch_size 8 --model FlowNet2C --optimizer=Adam --loss=MultiScale --loss_norm=L1 --loss_numScales=5 --loss_startScale=4 --optimizer_lr=1e-4 --crop_size 384 512 --training_dataset MpiSintelFinal --training_dataset_root mpi-sintel/training  --validation_dataset MpiSintelClean --validation_dataset_root mpi-sintel/training

Contributors

Developed by Cornelia Fermuller, Chethan Parameshwara, and Levi Burner with the Perception and Robotics Group at the University of Maryland.