{ "id": 388560, "title": "Two weeks left - Summary of the top discussions!", "url": "/competitions/nfl-player-contact-detection/discussion/388560", "author_name": "", "post_date": "2023-02-18T07:14:20.380024Z", "votes": 10, "comment_count": 1, "views": 0, "content": "

Two weeks left - Summary of the top discussions!

\n

We are approaching the end of the competition, here is a summary of all the top discussion posts so far.

\n
\n
The Matthews correlation coefficient (MCC) By Marília Prata
\n\n
\n
Evaluated on Matthews Correlation Coefficient Python [code] By Wongi Park
\n\n
\n
Matthews Correlation Coefficient[MCC] for Loss function [code]. It's can useful this competition. By Wongi Park
\n\n
\n
Articles to read for faster video data processing By Satya
\n\n
\n
Introduction to Tracking with SORT and DeepSORT By Ravi Shah
\n\n
\n
More Understanding of the data By Naif Alkhunaizi
\n\n
\n

Answer from the host Rob Mulla

\n

A few things to keep in mind about the two data sources:

\n

Tracking data is collected at 10Hz every 0.1 second should have a datapoint.
\n Video data is collected at 59.94Hz.
\n Submissions are at the tracking data 10Hz.
\n Because of the differences in sample rate it's impossible to perfectly line up the two datasets. One approach is to use the metadata file which gives the approximate start_time of each video, then calculate the approximate timestamp of each frame to merge with the tracking data.

\n

You can check the join_helmets_contact function in the starter notebook here: https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started

\n

You can also look at solutions from prior years to see how others approached merging this data.

\n

Hope that helps. Good luck!

\n
\n
\n
Should the model predict player Ids? By Ricard Fos
\n\n
\n

Answer from the host Rob Mulla

\n

Great question Richard. The goal of this competition is to predict when contact occurs between:
\n 1) Player pairs (any moment they are in physical contact)
\n 2) Player and the ground (When any part of the player's body except their feet or hands are touching the ground)

\n

You are asked to predict this for steps of 0.1 seconds within the play.

\n

The contact_id is simply a unique identifier for each moment you are expected to predict.

\n

It consists of either :

\n

{game_play}{step}{nfl_player_id_1}{nfl_player_id_2} (where nfl_player_id_1 < nfl_player_id_2) for player pair contact or\n{game_play}{step}_{nfl_player_id_1}_G for player to ground contact.
\n You will need to predict for every player pair and player/ground for each time step in the play (from step 0 to the last step in tracking or video data). Once submitted your notebook will have access to a sample_submission.csv which contains all of the contact_ids you are expected to predict for.
\n As for the data provided, you can use some or all of them to help you make predictions. They include:

\n

Tracking data which provides information about each player (position, speed, etc) at each step in the play.
\n Video of the play from three views (Sideline, Endzone and All29). Sideline and Endzone can be synced with the step. All29 is not guaranteed to be time synced.
\n Imperfect helmet boxes detected in the Sideline/Endzone videos which can be used or combined with the video to help your predictions.
\n How you choose to use the provided data is up to you. One of the things we are hoping to learn from this competition are the novel approaches kagglers take in using this data.

\n

Hope that helps!

\n
\n
\n
In game_play 58538_002774, there are 10 players in away team By NAMBA
\n\n
\n

Answer from the host Rob Mulla

\n

Good find,

\n

Initially, I suspected that this might be an issue with the tracking data. However, after looking into it, it appears that this is simply an unusual case where the defensive team only had 10 players on the field. This is very rare and not due to disqualification (teams are always permitted to have 11 players on the field during a play). It is possible that there was some confusion and the team unintentionally ended up with only 10 players on the field.

\n
\n

Hope that helps.

\n
\n
All training frame extracted dataset By Bao Loc Pham
\n\n
\n
Obviously wrong labels? By NxGTR
\n\n
\n

Answer from the host Rob Mulla

\n

Labeling a dataset of this size is a complex task, and while we took steps to ensure the accuracy of the labels, it is possible that some mistakes may still be present. These mistakes, if any, were not intentional, and we would consider them \"unknown\" rather than \"known\" mistakes.

\n

Regarding the large distance when contact occurs, there could be various reasons for this, including sensor delay and/or noise.

\n

We have no plans to alter or change the official data or labels, but if you come across any labels that you believe to be incorrect, please feel free to share them on the message board so that others can be made aware.

\n

Hope this helps to clarify things!

\n
\n
\n
How long does your submission take? By arutema47
\n\n
\n

Tracking vs Image data By Vladislav Ostankovich

\n\n
\n
Demosite - Albumentations By Ali Abdin
\n\n
\n

Good luck to you all in the final two weeks!

", "messages": [ { "id": "2149335", "postDate": "02/18/2023 07:14:20", "content": "

Two weeks left - Summary of the top discussions!

\n

We are approaching the end of the competition, here is a summary of all the top discussion posts so far.

\n
\n
The Matthews correlation coefficient (MCC) By Marília Prata
\n\n
\n
Evaluated on Matthews Correlation Coefficient Python [code] By Wongi Park
\n\n
\n
Matthews Correlation Coefficient[MCC] for Loss function [code]. It's can useful this competition. By Wongi Park
\n\n
\n
Articles to read for faster video data processing By Satya
\n\n
\n
Introduction to Tracking with SORT and DeepSORT By Ravi Shah
\n\n
\n
More Understanding of the data By Naif Alkhunaizi
\n\n
\n

Answer from the host Rob Mulla

\n

A few things to keep in mind about the two data sources:

\n

Tracking data is collected at 10Hz every 0.1 second should have a datapoint.
\n Video data is collected at 59.94Hz.
\n Submissions are at the tracking data 10Hz.
\n Because of the differences in sample rate it's impossible to perfectly line up the two datasets. One approach is to use the metadata file which gives the approximate start_time of each video, then calculate the approximate timestamp of each frame to merge with the tracking data.

\n

You can check the join_helmets_contact function in the starter notebook here: https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started

\n

You can also look at solutions from prior years to see how others approached merging this data.

\n

Hope that helps. Good luck!

\n
\n
\n
Should the model predict player Ids? By Ricard Fos
\n\n
\n

Answer from the host Rob Mulla

\n

Great question Richard. The goal of this competition is to predict when contact occurs between:
\n 1) Player pairs (any moment they are in physical contact)
\n 2) Player and the ground (When any part of the player's body except their feet or hands are touching the ground)

\n

You are asked to predict this for steps of 0.1 seconds within the play.

\n

The contact_id is simply a unique identifier for each moment you are expected to predict.

\n

It consists of either :

\n

{game_play}{step}{nfl_player_id_1}{nfl_player_id_2} (where nfl_player_id_1 < nfl_player_id_2) for player pair contact or\n{game_play}{step}_{nfl_player_id_1}_G for player to ground contact.
\n You will need to predict for every player pair and player/ground for each time step in the play (from step 0 to the last step in tracking or video data). Once submitted your notebook will have access to a sample_submission.csv which contains all of the contact_ids you are expected to predict for.
\n As for the data provided, you can use some or all of them to help you make predictions. They include:

\n

Tracking data which provides information about each player (position, speed, etc) at each step in the play.
\n Video of the play from three views (Sideline, Endzone and All29). Sideline and Endzone can be synced with the step. All29 is not guaranteed to be time synced.
\n Imperfect helmet boxes detected in the Sideline/Endzone videos which can be used or combined with the video to help your predictions.
\n How you choose to use the provided data is up to you. One of the things we are hoping to learn from this competition are the novel approaches kagglers take in using this data.

\n

Hope that helps!

\n
\n
\n
In game_play 58538_002774, there are 10 players in away team By NAMBA
\n\n
\n

Answer from the host Rob Mulla

\n

Good find,

\n

Initially, I suspected that this might be an issue with the tracking data. However, after looking into it, it appears that this is simply an unusual case where the defensive team only had 10 players on the field. This is very rare and not due to disqualification (teams are always permitted to have 11 players on the field during a play). It is possible that there was some confusion and the team unintentionally ended up with only 10 players on the field.

\n
\n

Hope that helps.

\n
\n
All training frame extracted dataset By Bao Loc Pham
\n\n
\n
Obviously wrong labels? By NxGTR
\n\n
\n

Answer from the host Rob Mulla

\n

Labeling a dataset of this size is a complex task, and while we took steps to ensure the accuracy of the labels, it is possible that some mistakes may still be present. These mistakes, if any, were not intentional, and we would consider them \"unknown\" rather than \"known\" mistakes.

\n

Regarding the large distance when contact occurs, there could be various reasons for this, including sensor delay and/or noise.

\n

We have no plans to alter or change the official data or labels, but if you come across any labels that you believe to be incorrect, please feel free to share them on the message board so that others can be made aware.

\n

Hope this helps to clarify things!

\n
\n
\n
How long does your submission take? By arutema47
\n\n
\n

Tracking vs Image data By Vladislav Ostankovich

\n\n
\n
Demosite - Albumentations By Ali Abdin
\n\n
\n

Good luck to you all in the final two weeks!

", "rawMarkdown": "### Two weeks left - Summary of the top discussions!\n\nWe are approaching the end of the competition, here is a summary of all the top discussion posts so far.\n\n______\n\n##### [The Matthews correlation coefficient (MCC)](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370688) By [Marília Prata](https://www.kaggle.com/mpwolke)\n\n- A good discussion topic by Marília Prata about the Matthews correlation coefficient metric and how to understand it.\n\n_____\n\n\n##### [Evaluated on Matthews Correlation Coefficient Python [code]](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370722) By [Wongi Park](https://www.kaggle.com/kalelpark)\n\n- Next, we got a first implementation of the metric made simple by [Wongi Park](https://www.kaggle.com/kalelpark)\n\n\n_____\n\n\n##### [Matthews Correlation Coefficient[MCC] for Loss function [code]. It's can useful this competition.](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370723) By [Wongi Park](https://www.kaggle.com/kalelpark)\n\n- We then got a version of the metric that can be used as a loss function (Really cool!)\n\n\n_____\n\n\n##### [Articles to read for faster video data processing](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371277) By [Satya](https://www.kaggle.com/satyaprakashshukl)\n\n- Some interesting articles about faster video data processing.\n\n_____\n\n\n##### [Introduction to Tracking with SORT and DeepSORT](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371702) By [Ravi Shah](https://www.kaggle.com/ravishah1)\n\n- This is a simple intro to DeepSort - The main model used in many previous solutions on the same subject.\n\n_____\n\n\n##### [More Understanding of the data](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371638) By [Naif Alkhunaizi](https://www.kaggle.com/naifalkhunaizi3)\n\n- [Naif Alkhunaizi](https://www.kaggle.com/naifalkhunaizi3) Was asking an important question about aligning the train labels with train_player_tracking and the video, since each video has approximately 711 frames, How do should we align these data with the video?\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> A few things to keep in mind about the two data sources:\n> \n> Tracking data is collected at 10Hz every 0.1 second should have a datapoint.\n> Video data is collected at 59.94Hz.\n> Submissions are at the tracking data 10Hz.\n> Because of the differences in sample rate it's impossible to perfectly line up the two datasets. One approach is to use the metadata file which gives the approximate start_time of each video, then calculate the approximate timestamp of each frame to merge with the tracking data.\n> \n> You can check the join_helmets_contact function in the starter notebook here: https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started\n> \n> You can also look at solutions from prior years to see how others approached merging this data.\n> \n> Hope that helps. Good luck!\n\n\n_____\n\n\n##### [Should the model predict player Ids?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371948) By [Ricard Fos](https://www.kaggle.com/ricardfos)\n\n\n- [Ricard Fos](https://www.kaggle.com/ricardfos) Was asking an important question about this competition: If the inputs will be just the contact_ids, which include the 2 player ids, game, play and step. we don't input the video?\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Great question Richard. The goal of this competition is to predict when contact occurs between:\n> 1) Player pairs (any moment they are in physical contact)\n> 2) Player and the ground (When any part of the player's body except their feet or hands are touching the ground)\n> \n> You are asked to predict this for steps of 0.1 seconds within the play.\n> \n> The contact_id is simply a unique identifier for each moment you are expected to predict.\n> \n> It consists of either :\n> \n> {game_play}_{step}_{nfl_player_id_1}_{nfl_player_id_2} (where nfl_player_id_1 < nfl_player_id_2) for player pair contact or\n> {game_play}_{step}_{nfl_player_id_1}_G for player to ground contact.\n> You will need to predict for every player pair and player/ground for each time step in the play (from step 0 to the last step in tracking or video data). Once submitted your notebook will have access to a sample_submission.csv which contains all of the contact_ids you are expected to predict for.\n> As for the data provided, you can use some or all of them to help you make predictions. They include:\n> \n> Tracking data which provides information about each player (position, speed, etc) at each step in the play.\n> Video of the play from three views (Sideline, Endzone and All29). Sideline and Endzone can be synced with the step. All29 is not guaranteed to be time synced.\n> Imperfect helmet boxes detected in the Sideline/Endzone videos which can be used or combined with the video to help your predictions.\n> How you choose to use the provided data is up to you. One of the things we are hoping to learn from this competition are the novel approaches kagglers take in using this data.\n> \n> Hope that helps!\n\n\n_____\n\n\n##### [In game_play 58538_002774, there are 10 players in away team](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/372241) By [NAMBA](https://www.kaggle.com/yasumasanamba)\n\n\n- [NAMBA](https://www.kaggle.com/yasumasanamba) has found out that In game_play 58538_002774, there are 10 players in away team, is this due to a player disqualification?\n\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Good find,\n> \n> Initially, I suspected that this might be an issue with the tracking data. However, after looking into it, it appears that this is simply an unusual case where the defensive team only had 10 players on the field. This is very rare and not due to disqualification (teams are always permitted to have 11 players on the field during a play). It is possible that there was some confusion and the team unintentionally ended up with only 10 players on the field.\n\nHope that helps.\n\n\n_____\n\n\n##### [All training frame extracted dataset](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/374646) By [Bao Loc Pham](https://www.kaggle.com/locbaop)\n\n\n- [Bao Loc Pham](https://www.kaggle.com/locbaop) had created [a dataset](https://www.kaggle.com/datasets/locbaop/nfl-contact-extracted-train-frames)\n- Extracted from [zzy990106](https://www.kaggle.com/zzy990106) [notebook](https://www.kaggle.com/code/zzy990106/nfl-2-5d-cnn-baseline-inference/notebook)\n- Total images: ~373k files\n\n\n_____\n\n\n##### [Obviously wrong labels?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/372806) By [NxGTR](https://www.kaggle.com/carloshuertas)\n\n- [NxGTR](https://www.kaggle.com/carloshuertas) has raised the fact that some labels look plain wrong, e.g. some players touching each other at 5yard distance.\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Labeling a dataset of this size is a complex task, and while we took steps to ensure the accuracy of the labels, it is possible that some mistakes may still be present. These mistakes, if any, were not intentional, and we would consider them \"unknown\" rather than \"known\" mistakes.\n> \n> Regarding the large distance when contact occurs, there could be various reasons for this, including sensor delay and/or noise.\n> \n> We have no plans to alter or change the official data or labels, but if you come across any labels that you believe to be incorrect, please feel free to share them on the message board so that others can be made aware.\n> \n> Hope this helps to clarify things!\n\n\n_____\n\n\n##### [How long does your submission take?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/379265) By [arutema47](https://www.kaggle.com/kyoshioka47)\n\n- An interesting post comparing submission times of many competitors using multiple different solutions.\n\n\n\n_____\n\n\n#### [Tracking vs Image data](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/379434) By [Vladislav Ostankovich](https://www.kaggle.com/vostankovich)\n\n- An interesting discussion about the information that can be extracted from the videos vs the information that can be extracted from the other metadata (With actual submission scores reports).\n\n\n____\n\n\n##### [Demosite - Albumentations](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/383841) By [Ali Abdin](https://www.kaggle.com/aliabdin1)\n\n\n- A simple [interface](https://demo.albumentations.ai/) for playing around with the different augmentations in the albuminations library. \n\n\n_____\n\nGood luck to you all in the final two weeks!", "votes": null }, { "id": "2241458", "postDate": "05/01/2023 14:18:34", "content": "

The significant thing when other Kagglers mention us, is that the chance of being doomed by oblivion decreases.

\n

Women complain frequently that they've been ever erased in so many fields. However, they should participate more if they really want to avoid that \"curse\" (misfortune).

\n

Thank you for mentioning my (Matthews correlation coefficient metric) Topic among so brilliant contributions. Probably, it was the first time I read something about the MCC.

\n

Writing and bringing material is a great way to learn with the others.

\n

Thank you The Devastator.

\n

Kind regards,
\nMarília Prata.

", "rawMarkdown": "The significant thing when other Kagglers mention us, is that the chance of being doomed by oblivion decreases.\n\nWomen complain frequently that they've been ever erased in so many fields. However, they should participate more if they really want to avoid that \"curse\" (misfortune).\n\nThank you for mentioning my (Matthews correlation coefficient metric) Topic among so brilliant contributions. Probably, it was the first time I read something about the MCC. \n\nWriting and bringing material is a great way to learn with the others.\n\nThank you The Devastator.\n\nKind regards,\nMarília Prata.", "votes": null } ], "comments": [ { "id": 2241458, "author_name": "mpwolke", "author_url": "", "post_date": "05/01/2023 14:18:34", "content": "

The significant thing when other Kagglers mention us, is that the chance of being doomed by oblivion decreases.

\n

Women complain frequently that they've been ever erased in so many fields. However, they should participate more if they really want to avoid that \"curse\" (misfortune).

\n

Thank you for mentioning my (Matthews correlation coefficient metric) Topic among so brilliant contributions. Probably, it was the first time I read something about the MCC.

\n

Writing and bringing material is a great way to learn with the others.

\n

Thank you The Devastator.

\n

Kind regards,
\nMarília Prata.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2149335": "### Two weeks left - Summary of the top discussions!\n\nWe are approaching the end of the competition, here is a summary of all the top discussion posts so far.\n\n______\n\n##### [The Matthews correlation coefficient (MCC)](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370688) By [Marília Prata](https://www.kaggle.com/mpwolke)\n\n- A good discussion topic by Marília Prata about the Matthews correlation coefficient metric and how to understand it.\n\n_____\n\n\n##### [Evaluated on Matthews Correlation Coefficient Python [code]](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370722) By [Wongi Park](https://www.kaggle.com/kalelpark)\n\n- Next, we got a first implementation of the metric made simple by [Wongi Park](https://www.kaggle.com/kalelpark)\n\n\n_____\n\n\n##### [Matthews Correlation Coefficient[MCC] for Loss function [code]. It's can useful this competition.](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/370723) By [Wongi Park](https://www.kaggle.com/kalelpark)\n\n- We then got a version of the metric that can be used as a loss function (Really cool!)\n\n\n_____\n\n\n##### [Articles to read for faster video data processing](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371277) By [Satya](https://www.kaggle.com/satyaprakashshukl)\n\n- Some interesting articles about faster video data processing.\n\n_____\n\n\n##### [Introduction to Tracking with SORT and DeepSORT](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371702) By [Ravi Shah](https://www.kaggle.com/ravishah1)\n\n- This is a simple intro to DeepSort - The main model used in many previous solutions on the same subject.\n\n_____\n\n\n##### [More Understanding of the data](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371638) By [Naif Alkhunaizi](https://www.kaggle.com/naifalkhunaizi3)\n\n- [Naif Alkhunaizi](https://www.kaggle.com/naifalkhunaizi3) Was asking an important question about aligning the train labels with train_player_tracking and the video, since each video has approximately 711 frames, How do should we align these data with the video?\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> A few things to keep in mind about the two data sources:\n> \n> Tracking data is collected at 10Hz every 0.1 second should have a datapoint.\n> Video data is collected at 59.94Hz.\n> Submissions are at the tracking data 10Hz.\n> Because of the differences in sample rate it's impossible to perfectly line up the two datasets. One approach is to use the metadata file which gives the approximate start_time of each video, then calculate the approximate timestamp of each frame to merge with the tracking data.\n> \n> You can check the join_helmets_contact function in the starter notebook here: https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started\n> \n> You can also look at solutions from prior years to see how others approached merging this data.\n> \n> Hope that helps. Good luck!\n\n\n_____\n\n\n##### [Should the model predict player Ids?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/371948) By [Ricard Fos](https://www.kaggle.com/ricardfos)\n\n\n- [Ricard Fos](https://www.kaggle.com/ricardfos) Was asking an important question about this competition: If the inputs will be just the contact_ids, which include the 2 player ids, game, play and step. we don't input the video?\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Great question Richard. The goal of this competition is to predict when contact occurs between:\n> 1) Player pairs (any moment they are in physical contact)\n> 2) Player and the ground (When any part of the player's body except their feet or hands are touching the ground)\n> \n> You are asked to predict this for steps of 0.1 seconds within the play.\n> \n> The contact_id is simply a unique identifier for each moment you are expected to predict.\n> \n> It consists of either :\n> \n> {game_play}_{step}_{nfl_player_id_1}_{nfl_player_id_2} (where nfl_player_id_1 < nfl_player_id_2) for player pair contact or\n> {game_play}_{step}_{nfl_player_id_1}_G for player to ground contact.\n> You will need to predict for every player pair and player/ground for each time step in the play (from step 0 to the last step in tracking or video data). Once submitted your notebook will have access to a sample_submission.csv which contains all of the contact_ids you are expected to predict for.\n> As for the data provided, you can use some or all of them to help you make predictions. They include:\n> \n> Tracking data which provides information about each player (position, speed, etc) at each step in the play.\n> Video of the play from three views (Sideline, Endzone and All29). Sideline and Endzone can be synced with the step. All29 is not guaranteed to be time synced.\n> Imperfect helmet boxes detected in the Sideline/Endzone videos which can be used or combined with the video to help your predictions.\n> How you choose to use the provided data is up to you. One of the things we are hoping to learn from this competition are the novel approaches kagglers take in using this data.\n> \n> Hope that helps!\n\n\n_____\n\n\n##### [In game_play 58538_002774, there are 10 players in away team](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/372241) By [NAMBA](https://www.kaggle.com/yasumasanamba)\n\n\n- [NAMBA](https://www.kaggle.com/yasumasanamba) has found out that In game_play 58538_002774, there are 10 players in away team, is this due to a player disqualification?\n\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Good find,\n> \n> Initially, I suspected that this might be an issue with the tracking data. However, after looking into it, it appears that this is simply an unusual case where the defensive team only had 10 players on the field. This is very rare and not due to disqualification (teams are always permitted to have 11 players on the field during a play). It is possible that there was some confusion and the team unintentionally ended up with only 10 players on the field.\n\nHope that helps.\n\n\n_____\n\n\n##### [All training frame extracted dataset](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/374646) By [Bao Loc Pham](https://www.kaggle.com/locbaop)\n\n\n- [Bao Loc Pham](https://www.kaggle.com/locbaop) had created [a dataset](https://www.kaggle.com/datasets/locbaop/nfl-contact-extracted-train-frames)\n- Extracted from [zzy990106](https://www.kaggle.com/zzy990106) [notebook](https://www.kaggle.com/code/zzy990106/nfl-2-5d-cnn-baseline-inference/notebook)\n- Total images: ~373k files\n\n\n_____\n\n\n##### [Obviously wrong labels?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/372806) By [NxGTR](https://www.kaggle.com/carloshuertas)\n\n- [NxGTR](https://www.kaggle.com/carloshuertas) has raised the fact that some labels look plain wrong, e.g. some players touching each other at 5yard distance.\n\n> **Answer from the host [Rob Mulla](https://www.kaggle.com/robikscube)**\n> \n> Labeling a dataset of this size is a complex task, and while we took steps to ensure the accuracy of the labels, it is possible that some mistakes may still be present. These mistakes, if any, were not intentional, and we would consider them \"unknown\" rather than \"known\" mistakes.\n> \n> Regarding the large distance when contact occurs, there could be various reasons for this, including sensor delay and/or noise.\n> \n> We have no plans to alter or change the official data or labels, but if you come across any labels that you believe to be incorrect, please feel free to share them on the message board so that others can be made aware.\n> \n> Hope this helps to clarify things!\n\n\n_____\n\n\n##### [How long does your submission take?](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/379265) By [arutema47](https://www.kaggle.com/kyoshioka47)\n\n- An interesting post comparing submission times of many competitors using multiple different solutions.\n\n\n\n_____\n\n\n#### [Tracking vs Image data](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/379434) By [Vladislav Ostankovich](https://www.kaggle.com/vostankovich)\n\n- An interesting discussion about the information that can be extracted from the videos vs the information that can be extracted from the other metadata (With actual submission scores reports).\n\n\n____\n\n\n##### [Demosite - Albumentations](https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/383841) By [Ali Abdin](https://www.kaggle.com/aliabdin1)\n\n\n- A simple [interface](https://demo.albumentations.ai/) for playing around with the different augmentations in the albuminations library. \n\n\n_____\n\nGood luck to you all in the final two weeks!", "2241458": "The significant thing when other Kagglers mention us, is that the chance of being doomed by oblivion decreases.\n\nWomen complain frequently that they've been ever erased in so many fields. However, they should participate more if they really want to avoid that \"curse\" (misfortune).\n\nThank you for mentioning my (Matthews correlation coefficient metric) Topic among so brilliant contributions. Probably, it was the first time I read something about the MCC. \n\nWriting and bringing material is a great way to learn with the others.\n\nThank you The Devastator.\n\nKind regards,\nMarília Prata." }, "source": "meta" }