Kynisys Challenge #1

Emotion Recognition

Kynisys is excited to host our 1st Edge AI competition. Emotion recognition has attracted a lot of attention in recent years due to a wide range of edge applications.

We’re looking for people who want to create a solution and can correctly predict the emotion class for an image within the limitations of edge devices.

What you'll get

Jupyter Notebooks:
Tensorflow 2.1 & Python 3.7

CAER-S Dataset: 70,000
annotated images


Train on GPU’s

Prize money:

Test on Raspberry Pi

Competition details

Kynisys is hosting an exciting challenge: Emotion Detection in the Wild. We believe this will help build innovative new technologies for the future.

The goal

Create a model that can correctly predict the emotion class for an image, within the limitations of edge devices. Here are some more details to take into consideration when building your models.

Bullet Point - CircleCreated with Sketch.

Build models ensuring the right balance between speed, accuracy and resource consumption

Bullet Point - CircleCreated with Sketch.

Build solutions using Tensor 2.1

Bullet Point - CircleCreated with Sketch.

Build solutions by coding in Jupyter notebooks environment on our platform and use free compute resources in the cloud and device farm

Bullet Point - CircleCreated with Sketch.

Target edge devices are Raspberry Pi 3b+

Bullet Point - CircleCreated with Sketch.

You decide how to use the supplied training subset, including the split for training and validation

Bullet Point - CircleCreated with Sketch.

Test dataset will be fixed and will be used to compute the public leaderboards

Available dataset

Dataset for training and testing is supplied by Kynisys.

Bullet Point - CircleCreated with Sketch.

The competitors will get access to the CAER-S Dataset: Context-Aware Emotion Recognition. It’s a large-scale static emotion recognition benchmark that contains more than 70,000 annotated images. The benchmark consists of train, validation and test folders.

The prizes

Kynisys is offering a total prize pot of $10,000 to be won.


1st prize



2nd prize



3rd prize



4th prize


The Kynisys Platform

The platform allows competitors to code/build their solutions using Jupyter notebooks.

Supported framework is TensorFlow 2.1 on python 3.7

Competitors have unlimited CPU/GPU resources, however, GPU use may be throttled in the case of many concurrent competitors

Jupyter notebooks are running on CPU resources (no quota limitation) by default, these can be used to code and debug the solution

Once complete, a notebook can be run in a captive flow that trains on a GPU cluster and evaluates the test both on the training system and the Raspberry Pi edge device

The conversion to the edge device uses TensorFlow Lite post-training quantisation with the representative “dataset”

Emotion recognition use cases

Exciting Emotion Recognition use cases are emerging across a set of industries.


Prevention of terrorism, smuggling, aggression, modern slavery.

Detecting the mood or possible intentions of people in any public area enables alerts of suspicious activity to be flagged to security staff in real time.


Medical diagnosis, help autistic people understand emotions.

Using emotion detection to interpret genuine pain and discomfort including location, speeding up diagnosis and filtering out bogus cases.


Prevent fatigue &

The immediate detection of negative driver emotions (anger, fatigue, stress, confusion, etc.) or unwanted distractions can give a very early warning of danger or potential issues.

Competition timeline

The competition will run towards the end of April.

Register your interest here.

24th April

Pre-registration open

18th May

Official entry open

31st May

Final submission deadline

15th June

Winners announced

Evaluation criteria

Bullet Point - CircleCreated with Sketch.

Multiple solutions per competitor can be submitted via Kynisys leaderboard

Bullet Point - CircleCreated with Sketch.

Competitor chooses which solution/run they submit

Bullet Point - CircleCreated with Sketch.

Public leaderboard will display scores obtained by running solutions on the test subset from the dataset

Bullet Point - CircleCreated with Sketch.

Score will be based on top-1 accuracy of the model, however, submissions will only be possible if the constraints on inference time (100ms) have been met

Interested to pre-register your interest?