Ant societies have long fascinated mankind by their resemblance to human societies and elaborate collective behavior. Ants cultivate and farm fungus in their nests, herd aphids as cattle and wage war with each other.
In contrast to human societies, ants present an ideal study system to study how collective patterns emerge through self-organization and how societies respond to perturbations in their environment. We can easily manipulate crowding, resource availability and parasite pressure, thereby allowing us to study questions that would be impossible to answer in humans. One of the most interesting collective dynamics in a colony is the trade-off between fast transmission of food and how ants avoid the outbreak of diseases.
To better understand transmission dynamics and the underlying socio-ecological drivers, we have to analyze ant behaviour across time and space. The goal of this challenge is to generate algorithms that can:
In the first part of this challenge, we will focus on the task of identification and tracking of individual ants over time. The training data provides the coordinates of a subset of the ants for all the time frames, and the goal of the challenge is to predict the coordinates of the rest of the ants for all the time frames.
To identify individuals in the colony, we have attached a unique barcode (as described in http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0136487 Crall JD, Gravish N, Mountcastle AM, Combes SA (2015) BEEtag: A Low-Cost, Image-Based Tracking System for the Study of Animal Behavior and Locomotion. PLoS ONE 10(9): e0136487. doi:10.1371/journal.pone.0136487 ) to the gaster of each worker (see image 1).
The only individual that was not marked was the queen, since her distinct appearance (larger body size) makes her very easy to identify, and moreover, we did not want to stress or injure her. Students have manually tracked several hours of ant movement inside a colony allowing us to determine the location of each worker for each second of video footage.
To ensure that the students’ tracking data is as exact and repeatable as possible, we chose to track the neck area of each ant (see image 2) using customized Python code. We were able to obtain x and y coordinates (unit is pixels) for the location of all individuals at each second of the observation period. In instances when the ant neck was not visible, e.g., when an ant was underneath another ant, we estimated the neck position based on the last known location and the other visible body parts of this ant. If an ant left the nest, its coordinates were set to x = 0, y = 0.
Challenge Logo Source : https://www.flickr.com/photos/98180998@N04/9187248340/
The training data provides the coordinates of a subset of the ants for all the time frames, and the goal of the challenge is to predict the coordinates of the rest of the ants for all the time frames.
The overall error (or Loss) functions are defined as :
Note, all boolean variables will be treated as integral 1 when they are True and 0 when they are false, for the computations in L_1.
The goal of the challenge is to minimise this error function/loss L_1 (which approximates the overall accuracy), and if multiple participants have the same L_1 score, then they will be evaluated and ranked based on L_2 (which is the mean squared error).
All submissions will be evaluated on the test dataset in the docker containers referenced in the Resources section. The code archive will be uncompressed into the
/ants path, and every code archive is expected to contain a
main.sh script which takes path to a folder containing frame images that will be tested, as its first parameter. So to test your code submission, we will finally execute :
This is expected to output a CSV file containing the name of the file, and the associated probabilities for all the classes at the location :
The author of the most highly ranked submission will be invited to the Applied Machine Learning Days symposium at EPFL in Switzerland on January 30/31, 2017. This educational award is granted to the participant with the either the most insightful submission posts, or the best tutorial. Expenses for travel and accommodation are covered by crowdAI.
The submission needs to have at least an L_1 score of 80% to be eligible for this prize.
There is also a monetary prize for the most highly ranked submission, which is determined as follows :
$500 if the L_1 score of the best submission reaches >= 80%, but less than 85%
$1000 if the L_1 score of the best submission reaches >= 85%, but less than 90%
$2000 if the L_1 score of the best submission reaches >= 90%, but less than 95%
$5000 if the L_1 score of the best submission reaches >= 95%
The challenge dataset contains 6 files:
cat frames.tar.gz.part-1 frames.tar.gz.part-2 frames.tar.gz.part-3 > frames.tar.gz
The code must be runnable in one of these Docker containers:
Caffe : https://hub.docker.com/r/tleyden5iwx/caffe-gpu-master/
Tensorflow : https://hub.docker.com/r/tensorflow/tensorflow/
Torch7 : https://hub.docker.com/r/kaixhin/cuda-torch/
Scikit-Learn :(Python-2): https://github.com/dataquestio/ds-containers/tree/master/python2
Scikit-Learn : (Python-3): https://github.com/dataquestio/ds-containers/tree/master/python3
Octave : https://hub.docker.com/r/schickling/octave/
Keras : https://hub.docker.com/r/patdiscvrd/keras/~/dockerfile/
If you wish to use another coding environment please contact here.
For submission, leaderboard and other details, visit <https://www.crowdai.org/challenges/ants-challenge-part-1>
Brought to you by: