sync.soccer synchronises (or aligns) soccer datasets, so that an event in one dataset is matched to the corresponding event or snapshot in the other. This repository contains an implementation that aligns Opta's (now STATS Perform's) F24 feeds to ChyronHego's Tracab files. More formats may be added in the future. The heart of sync.soccer is the Needleman-Wunsch dynamic programming algorithm. For a fuller explanation of the method see separate blogpost.
sync.soccer has been developed and tested on Ubuntu Linux. Feedback on compilation and performance on other platforms is welcome. To build sync.soccer, you need Stack. Then, to download and compile, say:
$ git clone [email protected]:huffyhenry/sync.soccer.git
$ cd sync.soccer
$ stack buildIf the build was successful, you can run sync.soccer like so (note the double dash, which tells Stack that all arguments that follow are meant for the sync-soccer executable and not for Stack itself):
$ stack exec -- sync-soccer --help
Usage: sync-soccer TRACAB-META TRACAB-DATA F7 F24 OUTPUT [-t|--time-only]
[-u|--timestamp] [-s|--show-sync] [-c|--scale-clock X]
[-l|--scale-location X] [-p|--scale-player X]
[-b|--scale-ball X] [-e|--event-csv FILEPATH]
[-f|--frame-csv FILEPATH]
Synchronise Tracab and F24 data.
Available options:
-t,--time-only Sync only by time
-u,--timestamp Use F24 event timestamp instead of min:sec
-s,--show-sync Print human-readable sync on screen
-c,--scale-clock X Clock mismatch leading to unit penalty [s, def=1]
-l,--scale-location X Location mismatch leading to unit penalty [m, def=5]
-p,--scale-player X Player-ball gap leading to unit penalty [m, def=1]
-b,--scale-ball X Penalty for syncing to dead-ball frame [def=5]
-e,--event-csv FILEPATH Location to save F24 events CSV
-f,--frame-csv FILEPATH Location to save Tracab frames CSV
-h,--help Show this help text
You must specify five file paths: to Tracab metadata, Tracab payload, F7 file (essentially F24 metadata), F24 file and a path to the output file where the completed alignment will be saved as a CSV of event and frame IDs.
The meaning of the optional arguments is as follows:
--time-onlydiscards all the added value of the algorithm and matches the events to the frame nearest to it in time. Useful for debugging, comparisons or when the algorithm misbehaves on a particuar game.--timestampuses the F24 event timestamp instead of minute and second as the time of the event.--show-syncwill print the alignment on screen. It will look something like this:Event <event id> (67:12 Pass) Frame <frame id> (implied clock: 67:12.920) --gap length 55-- Event <event id> (67:15 Pass) Frame <frame id> (implied clock: 67:15.160) --gap length 30-- Event <event id> (67:16 Clearance) Frame <frame id> (implied clock: 67:16.400) --gap length 26-- Event <event id> (67:17 Ball recovery) Frame <frame id> (implied clock: 67:17.480) Event <event id> (67:18 Pass) Frame <frame id> (implied clock: 67:17.520) --gap length 35-- Event <event id> (67:19 Pass) Frame <frame id> (implied clock: 67:18.960) --gap length 50-- Event <event id> (67:21 Pass) Frame <frame id> (implied clock: 67:21.000) --gap length 118-- Event <event id> (67:26 Pass) Frame <frame id> (implied clock: 67:25.760) --gap length 104-- Event <event id> (67:30 Pass) Frame <frame id> (implied clock: 67:29.960) --gap length 42-- Event <event id> (67:32 Miss) Frame <frame id> (implied clock: 67:31.680)- The parameters starting with
--scalecontrol the relative importance of the four components of the event-frame similarity score: difference of time, location, distance between the player identified by the event and the ball, and ball status (alive/dead). The default values were chosen by trial and error, and force the algorithm to penalise equally a 5 metre discrepancy in location, 1 second discrepancy in time, and 1 metre distance between the active player and the ball. Because the similarity score between any event and any frame falls in (0, 1), the default--scale-ballvalue of 5 means that the algorithm can never align to a dead-ball frame. --event-csvand--frame-csvwrite the event and frame data to files in a format compatible with the visualisation scripts inviz/.
You may have received a Tracab file without the corresponding metadata. Tracab metadata is a small XML file that nevertheless contains essential information for the synchronisation algorithm: pitch size, frame rate and the beginning and end of the first and second half. If you know these parameters or can deduce them from raw data, you can mock up a metadata file based on this template:
<TracabMetaData sVersion="1.0">
<match iId="0" dtDate="0" iFrameRateFps="25" fPitchXSizeMeters="105.00" fPitchYSizeMeters="68.00" fTrackingAreaXSizeMeters="0.00" fTrackingAreaYSizeMeters="0.00">
<period iId="1" iStartFrame="<value>" iEndFrame="<value>"/>
<period iId="2" iStartFrame="<value>" iEndFrame="<value>"/>
<period iId="3" iStartFrame="0" iEndFrame="0"/>
<period iId="4" iStartFrame="0" iEndFrame="0"/>
</match>
</TracabMetaData>Currently, sync.soccer cares only about the iFrameRateFps,
iStartFrame, iEndFrame, fPitchXSizeMeters and fPitchYSizeMeters
values. The rest can be set to 0. In fact, you only really need
the iStartFrame values: iEndFrame can safely be set to something
corresponding to a time after the half ended (at a small performace cost),
frame rate is usually (always?) 25, and you may get decent enough results
using the standard pitch size of 105x68 metres (and perhaps increasing the
--scale-location value).
The core algorithm needs to build two ~75'000 (# of Tracab frames) x ~800 (# of F24 events) matrices, one for each half of the match, to find the optimal alignment. This requires time and space. On my machine, the program needs 5-10 minutes and ~10 GB of RAM to synchronise a single game. There is a lot of low-hanging fruit to be picked in terms of optimisation, and I expect future versions of the software to be much more streamlined. Parallel synchronisation of game halves is also in the works.
Sometimes an alignment produced by sync.soccer contains segments that are demonstrably worse than if synchronisation were performed purely on the basis of time. They are typically a result of discrepancies between the F24 and Tracab data that are greater than is reasonable, for example due to very inaccurate F24 event location or even Tracab ball tracking error. In such situations, the algorithm may align event(s) to frame(s) that are a significant time apart, but where (by chance) the locations agree better. Such mis-alignment can then have a knock-on effect on the next several seconds of data, until everything straightens out.
One solution in these situations is to increase the --scale-location
parameter, making the algorithm less sensitive to differences in
event and ball locations. Another is to run a second synchronisation
with --time-only and manually (or using a script) merge the two sync
files so that the
problematic events are aligned only based on time. The time slippage
plot in viz/slippage.R can help identify the
potentially problematic segments, which can then be animated with
viz/animate.R to see what is going on.
Such pathological situations need to be distinguished from those when the algorithm produces an imperfect but reasonable alignment. These may eventually be corrected by developing smarter, more granular event-frame agreement scoring functions.
Copyright (c) 2018-2020 Allan Clark and Marek Kwiatkowski.
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.