{% extends 'base.html' %}

{% block title %}
Test Instructions{{ block.super }}
{% endblock %}

{% block content %}
    {{ block.super }}
    <div class='description'>
        <div class='panel'>
            <h3>Test Instructions</h3>
            <p>Given the (lat, lon) boundary points for the region, you must
                create a regular grid of 5001 x 5001 rectangular cells with the IPP at
                the center. In screen coordinates:
                <ul>
                    <li>The upper left corner is (0,0).</li>
                    <li>The upper right corner is (0,5000).</li>
                    <li>The bottom left corner is (5000, 0).</li>
                    <li>The bottom right corner is (5000, 5000).</li>
                    <li>The IPP is (2500,2500).</li>
                </ul>
                <strong class='error'>Note: We use a Great Sphere approximation of the Earth.</strong>
            </p>
            <p><strong>Each case supplies the 4 (lat,lon) points
                defining the corners of the search region, and the number of
                cells. Each cell must have the same height and width.</strong></p>
            <p><strong>Your model should rate each cell based on the
                probability that the lost person will be located in that
                cell. This process is left to you and your individual
                model. You are encouraged to use all provided test case
                information as well as all layers needed as input to your
                model.</strong></p>
            <p><strong>Ideally, your model works with probabilities. If
                so, the following will be true:
                <ul>
                    <li><i>P</i> + <i>Q</i> = 100%, where</li>
                    <li><i>P</i> = sum of probabilities for cells in the search region</li>
                    <li><i>Q</i> = probability subject is outside the search region</li>
                </ul>
                For now you must project these to <i>an 8-bit grayscale
                image</i> for display and scoring purposes. Pixel values range
                from 0 to 255, with 255 (white) being the highest relative value. (The
                scoring metric only considers rank order.) Our native format
                is grayscale PNG (color bands 'P' or 'L').</strong></p>
            <p><strong>If you work in probabilities, your numbers will
                span several orders of magnitude. There are 25 million pixels
                in the image, so a uniform map assigns each cell a probability
                of about 4x10<sup>-8</sup> (0.00000004). An actual map may
                have cell probabilities ranging several orders of magnitude
                above and below 10<sup>-8</sup>.The following suggested transform
                allows 17 distinctions at each order of magnitude from 1.0
                down to 10<sup>-15</sup> and even a few at
                10<sup>-16</sup>:
                <center>V = 5 * log<sub>2</sub>(P) + 255</center>
                In Python using numpy this would be:<br/>
                <center><code>V = 5 * int(numpy.log2(P)) + 255</code></center>
                If P were the matrix of probabilities, then V is now the
                matrix of grayscale values, though they may require converting
                to integers with <code>P.astype(int)</code> before saving.</strong></p>
            <p><strong>Lastly, you will be prompted to submit your
                grayscale image for analysis.  Once submitted, we will rate
                your image and "close" the test case for now. The new rating
                will be used to update your model's overall rating and your
                leaderboard status will be updated.</strong></p>
        </div>
    </div>
    <a href='/test/'>Return to test</a>&nbsp;|&nbsp;
    <a href='/model_menu/'>Return to model menu</a>
{% endblock %}