Data

Training data contains ~11500 examples, each of them consists of equivalently recorded 5-frame sequence.

In training data set, spoof class (fraud) has three separate categories:

• 2dmask - printed photo face mask;
• printed  - printed photo display;
• replay - video record replay.

Testing images are similar.

Spoof deems as a positive class of label=1. P(falsealarm) is a probability of a false alarm, i.e. real face classified as a spoof. P(miss) — the probability of missing a spoof.

Any open data sets and open pre-trained models are allowed for algorithm training.

Training data contains:

• Dataset (30 Gb) — full data set;
• Small train (10 Gb) — cropped data set;
• Local check data + sample submission — tiny data sample in testing environment format for local debugging.

Discussion and rules

Challenge discussion is held in the #idrnd_challenge channel of ODS.AI slack (registration is available here).

Teams are up to 6 people. Public code sharing is allowed strictly into the slack discussion channel.

Everyone can participate, except for LLC "DSP Labs" and "Singularis Lab" employees.

Challenge takes place from May 9 to June 20 inclusively. Team submission deadline - June 13, 23:59.

Solution format

The solution must include an archive containing the algorithm (code + necessary files) and startup point description. The root directory of a solution archive must include file meta.json of the following structure:

{
	"image": "<docker image>",
	"entrypoint": "<entry point or sh script>"
}

For example:

{
  "image": "ksanvatds/idrnd-antispoof",
  "entrypoint": "python3 predict.py --path-images-csv $PATH_INPUT/meta.csv --path-test-dir $PATH_INPUT --path-submission-csv $PATH_OUTPUT/solution.csv"
}

The solution runs in a docker-container with ksanvatds/idrnd-antispoof as a base image, similar to GPU image for Kaggle.

Organizers provide a correct baseline solution. Additional info is available at the GitHub page of a challenge.

Technical restrictions

• CPU - x4 2.3 GHz
• RAM - 8 Gb
• Solution size - up to 200 MB
• Work directory size - 4 GB
• Output directory size - 25 MB
• Time limit - 20 minutes
• GPU - NVIDIA Tesla K80

CUDA 10, CUDNN 7.4, and recent library versions are provided. If desired, own image could be dispensed.