1st Lacmus Computer Vision Competition

Project Lacmus

Every person going outdoors has a chance to get lost. Usually people get lost in the forest and wilderness. About 120 000 Persons are lost In Russia every year. More than 100 000 Persons are lost in the USA every year.

Usually a person, who is lost in wilderness, meets a number of risk factors, such as: Dehydration, Hypothermia, Trauma, Panic. Unfortunately, this person has a high chance of dying.

Various search and rescue teams around the world are looking for missing people. For searches, they often use unmanned aerial vehicles with which they perform aerial photography.

But... Analyzing photos also needs human resources. How can we speed up this process? We can use Neural Networks!

Our project, Lacmus, helps Search & Rescue Squads to find people, who get lost, by analyzing photos from copters with Neural Networks. And now we are ready for The First Competition!

At the competition, we offer you to solve the problem of recognizing and detecting lost people in aerial images. It is a classical one class object detection task. We need to detect people in photos from copters with an estimated altitude about 50m (164 ft).

Task Definition

During the competition, you are invited to solve the one class object detection task.

Metrics

We considering two metrics:

• mAP @ IoU 50
• F1 @ IoU 50

The implementation of these metrics can be found in the file metrics.py

Submit Format

The submission has csv format the following fields: id,xmin,ymin,xmax,ymax,score. Where:

• id - id of image (e.g. 1.jpg has id 1, 2.jpg has id 2 etc);
• xmin, ymin, xmax, ymax - bounding box coordinates;
• score - prediction confidence (float: [0..1])

The submission file must include a header.

Example:

id,xmin,ymin,xmax,ymax,score
0,127,2877,271,2921,0.589
0,1751,2111,1888,2222,0.111
1,583,1111,666,1211,0.799
...

ODS slack channel #ml4sg