The Bootcamp hosts MAFAT Satellite Vision Challenge! 🛰 [Online]

We are very happy to host MAFAT for a meetup launching their Satellite Vision Challenge!🛰

Can your Object Detection Model handle Model Drift?

We are excited to launch the “MAFAT Satellite Vision Challenge” — Satellite Imagery Object Detection Competition. This is the 4th competition in the MAFAT Challenge series. In this challenge, MAFAT would like to tackle the challenge of object detection from satellite imagery (see challenge details below).

Meetup Agenda:
18:30-18:50 - Idan Barak, VP Product at Webiks
Competition overview, starting from the registration process, and ending in collecting the prizes ($45,000 total prize pool!).

18:50-19:30 - Neta Mor & Shai Schneider, data scientists at Webiks
An in-depth explanation of the competition's unique dataset and a presentation of the baseline model will be provided for the competition participants.

19:30 - 20:15 - We are happy to host Nadav Barak, machine learning researcher at Deepchecks
drift detection in structured and unstructured data.

*All presentations will be given in English.

The meetup will be broadcasted via the Ministry of Defense YouTube channel - https://youtu.be/gxisjLWEV08
The challenge site is: https://mafatchallenge.mod.gov.il/

More about the challenge:
The challenge's participants goal is to detect and classify objects from various classes - such as Airplanes, Vehicles, Vessels, etc. - from diversified satellite images.

Participants will get access to a unique dataset containing thousands of labeled satellite images. The images differ by resolution (0.4m to 1.3m Ground Sample Distance), downward angle (nadir to off-nadir), direction (azimuth) and the date and time they were taken (time of day and seasonality).

Training great models on training data alone is not enough. Developing models that can adapt and maintain their performance in the face of changing input data and environments is a crucial task. In this challenge, for the first time in the series, we ask participants to tackle the problem of “model drift”!😱

The performance of ML models tend to degrade over time - as the model is deployed in a different context or environment than it was trained in. To deal with those drifts - participants’ models are allowed to have two passes on the test data - first one for calibration and second for predictions!🏆