Unified direction of play

[jan-swiatek]

Hi! I really appreciate work you put into this project :) I was wondering if there is any way to normalize direction of playing? Unfortunately different providers have different strategies. I'm interested mostly in normalizing statsbomb and tracab datasets to have consistent attacking direction. I couldn't find any guides in documentation. Thanks in advance, Jan :)

[UnravelSports]

Hi @jan-swiatek,

This is certainly possible, the simples way to do this is to use the .transform() functionality to call it on an EventDatasat and TrackingDataset, like so:

event_dataset = statsbomb.load(
    ....
    coordinates="secondspectrum"
).transform(
    to_orientation="STATIC_HOME_AWAY"
)

tracking_dataset = tracab.load(
    ....
    coordinates="secondspectrum"
).transform(
    to_orientation="STATIC_HOME_AWAY"
)

You can pass a coordinate system name to the loader, but you can also pass it to the transformer e.g. the below should give the same results:

event_dataset = statsbomb.load(
    ....
).transform(
    to_orientation="STATIC_HOME_AWAY",
    to_coordinate_system="secondspectrum"
)

You can opt to use any of the providers as the preferred coordinate system (e.g. "tracab", "secondspectrum", "statsbomb" etc.).

For orientation you can choose "STATIC_HOME_AWAY" (meaning home team is always left to right for the whole game) "STATIC_AWAY_HOME" (away team left to right for the whole game) and you can also use "BALL_OWNING_TEAM" or "ACTION_EXECUTING_TEAM" this will orient the attacking team from left to right. Additionally "HOME_AWAY" and "AWAY_HOME" will flip the orientation for each half.

There are even more settings for specifying your own pitch_dimensions too using the to_pitch_dimensions parameter in the transformer. Let me know if you need more info on that.

I'm using the secondspectrum coordinate system in this example because that's the one I prefer for most things.

For more info please see the "still under construction" docs that should be released soon/

[jan-swiatek]

Thank you @UnravelSports. I have one more question - since I have access to 360 data I would expect some additional columns like freeze_frame and visible_area after loading it, but I can't see them. How can I access this information?

[UnravelSports]

@jan-swiatek I assume you mean when converting using to_df()?

You can pass the to_df() stuff like this:

.to_df(
        "*",
        lambda event: {
            "is_pass": event.event_type == EventType.PASS,
            "freeze_frame": event.freeze_frame
        }
    )

Using the lambda's allows you to access each event and grab stuff from them that will be added as columns.

The above might not be 100% correct, but you can find some more examples on Enriching Event Data and Exporting to Dataframe in the docs

[jan-swiatek]

It more or less works :) Are freeze frames from 360 also being transformed? I mean if I apply transform with different coordinates system and orientation would that be applied to them as well?

[probberechts]

Yes, freeze frames are also being transformed.

If you want the freeze frames in a dataframe, it might be more convenient to first create a TrackingDataset from the freeze frames and then export the TrackingDataset as a dataframe.

[jan-swiatek]

Maybe I miss something, but for me visible area seems to not be transformed:

I also can not find players' coordinates in freeze frames. How to access their locations?

[probberechts]

Maybe I miss something, but for me visible area seems to not be transformed:

The visible_area is indeed not transformed. That's something you'll stil have to take care of yourself. Anything that is stored in .other_data is essentially ignored by kloppy. But as future work, it might be a good idea to add a visible_area attribute to the tracking data model.

I also can not find players' coordinates in freeze frames. How to access their locations?

The player coordinates are stored in the freeze_frame.players_coordinates attribute. You can take a look at the test file for an example of how to use it.

[jan-swiatek]

I see now. Due to long string representation I missed that this is a dictionary - I thought that this was set instead. It makes sense now, thanks :)

However, even though I applied .to_orientation=Orientation.HOME_AWAY I still sometimes can find consecutive freeze frames which seem to be mirrored. Did you experience some similar behavior?

[probberechts]

Can you provide an example where it fails using the public data?

[jan-swiatek]

Sure, I'll try to find an example over the weekend

[jan-swiatek]

In match 3895302 I found following examples:

Code to reproduce:

# %%
import numpy as np
import plotly.express as px

from kloppy import statsbomb
from kloppy.domain import Orientation
from statsbombpy import sb

# %%
match_ids = []

for _, competition in sb.competitions().iterrows():
    if competition["match_available_360"] is None:
        continue

    for _, game in sb.matches(competition["competition_id"], competition["season_id"]).iterrows():
        if game["match_status_360"] != "available":
            continue

        match_ids.append(game["match_id"])

# %%
statsbomb_dataset = statsbomb.load(
    event_data=f"https://raw.githubusercontent.com/statsbomb/open-data/master/data/events/{match_ids[0]}.json",
    lineup_data=f"https://raw.githubusercontent.com/statsbomb/open-data/master/data/lineups/{match_ids[0]}.json",
    three_sixty_data=f"https://raw.githubusercontent.com/statsbomb/open-data/master/data/three-sixty/{match_ids[0]}.json"
).transform(to_orientation=Orientation.HOME_AWAY)

# %%
counter = 0

for event in statsbomb_dataset.events:
    if event.freeze_frame is None:
        continue

    counter += 1

    print(event.event_id)

    points = np.array([[point.x, point.y] for point in event.freeze_frame.players_coordinates.values()] + [[event.coordinates.x, event.coordinates.y]])
    colors = np.array([player.team.ground for player in event.freeze_frame.players_coordinates.keys()] + ["ball"])

    fig = px.scatter(x=points[:, 0], y=points[:, 1], color=colors)
    
    fig.update_layout(width=700, height=400, title=f"{event.period.id} {event.timestamp} {event.team} {event.event_type} {event.event_name}")
    fig.update_xaxes(range=[0, 1])
    fig.update_yaxes(range=[0, 1])
    fig.show()

    if counter > 100:
        break

[UnravelSports]

Yes, freeze frames are also being transformed.

@probberechts any chance you could verify on one of your local installations, but the line you point to is incorrectly indented in all my local 3.17 installations. It is indented to only execute under the if self._needs_orientation_change condition, but in the link you shared it seems to not be the case.

[probberechts]

I incorrectly indented it in 3bca4f6. @BirCWld found it and was going to create a fix.