Question: Dataset

[daehoumich]

Hi, Agrim

I am trying to make my own dataset following the same format that you used (i.e., eth, zara, hotel, and univ).

The data files have four columns though, I couldn't clearly understand what the value of each column represents.

If you may, could you explain it?

[angeliand]

Hi daehoumich, Although I'm not Agrim, I hope my answer will be fit for your liking.

If you check out the sgan.data.trajectories.py file, and inside that the TrajectoryDataset class, you will find a docstring with the answer to your question:

- data_dir: Directory containing dataset files in the format
        <frame_id> <ped_id> <x> <y>

Just to make things a bit clearer:

• Every pedestrian has a unique ID (ped_id - second column), so that we can follow their movements/trajectories across different frames.
• Frames also have an ID (frame_id - first column), because we need to know the correct order of them.
• And finally, to know the position of a pedestrian x-y coordinates are given in the third and fourth column respectively.

[daehoumich]

Thank you so much~!

[mikegillimage]

Hi, this may sound silly, but does anyone know what is physical scale of the images in eth, zara and hotel datasets? Is there a way to get it from the homography matrix?

[amiryanj]

Hi, this may sound silly, but does anyone know what is physical scale of the images in eth, zara and hotel datasets? Is there a way to get it from the homography matrix?

I am just telling you the basic things. Having Homography matrix and a pixel coordinate [u,v], you should calculate H * [u, v, 1]^t => (x, y, z) and divide the resultant vector by (z) and you have the world coordinate: (x', y', 1). To know the boundaries of those scenes, it's enough to test the boundaries of the images; e.g. [0,0], [640, 480] for ETH!

[mikegillimage]

Thanks Amiryanj