{
  "id": 67089,
  "title": "Competition winners and Judges' Award",
  "url": "/competitions/freesound-audio-tagging/discussion/67089",
  "author_name": "Frederic Font",
  "post_date": "2018-09-28T16:21:45.035000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n\n<p>This post is to officially announce the winners of the <strong>Freesound General-Purpose Audio Tagging Challenge</strong> and the Judges' Award. </p>\n\n<h3>Challenge winners</h3>\n\n<p>These are the official winners of the competition and their score in the private leaderboard:</p>\n\n<ol>\n<li><strong>Cochlear.ai</strong> - 0.953810</li>\n<li><strong>Matthias Dorfer CPJKU</strong> - 0.951757</li>\n<li><strong>Surrey CVSSP</strong> - 0.951244</li>\n</ol>\n\n<p>Congratulations!</p>\n\n<p>One of the requirements for the winning teams was to open-source the code for the submission. Here is the link to the code repositories for the winning teams and the technical report submitted to the DCASE challenge:</p>\n\n<ul>\n<li>Cochlear.ai - <a href=\"https://github.com/finejuly/dcase2018_task2_cochlearai\">https://github.com/finejuly/dcase2018_task2_cochlearai</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Jeong_102.pdf\">technical report</a></li>\n<li>Matthias Dorfer CPJKU - <a href=\"https://github.com/CPJKU/dcase_task2\">https://github.com/CPJKU/dcase_task2</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Dorfer_999.pdf\">technical report</a></li>\n<li>Surrey CVSSP - <a href=\"https://github.com/turab95/dcase2018_task2\">https://github.com/turab95/dcase2018_task2</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Iqbal_89.pdf\">technical report</a></li>\n</ul>\n\n<p>We encourage you to visit the <a href=\"http://dcase.community/challenge2018/task-general-purpose-audio-tagging-results\">DCASE Task 2 results page</a> which includes an interesting summary of the characteristics of the systems of the teams who submitted both to Kaggle and the DCASE challenge.</p>\n\n<h3>Judges' award</h3>\n\n<p>After careful inspection of the challenge submissions who met the <a href=\"https://www.kaggle.com/c/freesound-audio-tagging/discussion/59932\">eligibility requirements</a> for the Judges' Award, we chose to award the following submission:</p>\n\n<ul>\n<li><strong>Emilia_NTU</strong> : The report includes an interesting and well-explained approach for dealing with the problem of potentially noisy labels in the non-verified portion of FSDKaggle2018. This self-supervision approach is based on pseudo-labeling and is complemented with label smoothing. Despite model ensembling was utilized (which was not preferred for the Judges' Award), competitive single-models have been considered too as a result of the self-supervision methodology. We’ve valued the insights disclosed from the systematic analysis conducted rather than the scores achieved. Also, the architecture proposed combines the internal representations learned with a CNN, and with few fully connected layers, making the most out of the input log-mel spectrogram. Last but not least we’ve valued the report readability.</li>\n</ul>\n\n<hr>\n\n<p>Aaaaand that's it for the Freesound General-Purpose Audio Tagging Challenge. Congratulations to all participants and specially to the winners! We hope you enjoyed participating as much as we did organizing it. We hope to come back next year with a bigger dataset and an even more challenging competition.</p>\n\n<p>Sincerely,</p>\n\n<p>Eduardo Fonseca, Frederic Font (Freesound, MTG-UPF)</p>\n\n<p>Dan Ellis, Manoj Plakal (Google Machine Perception)</p>",
  "messages": [
    {
      "id": 395467,
      "postDate": "2018-09-28T16:21:45.037Z",
      "content": "<p>Hi everyone,</p>\n\n<p>This post is to officially announce the winners of the <strong>Freesound General-Purpose Audio Tagging Challenge</strong> and the Judges' Award. </p>\n\n<h3>Challenge winners</h3>\n\n<p>These are the official winners of the competition and their score in the private leaderboard:</p>\n\n<ol>\n<li><strong>Cochlear.ai</strong> - 0.953810</li>\n<li><strong>Matthias Dorfer CPJKU</strong> - 0.951757</li>\n<li><strong>Surrey CVSSP</strong> - 0.951244</li>\n</ol>\n\n<p>Congratulations!</p>\n\n<p>One of the requirements for the winning teams was to open-source the code for the submission. Here is the link to the code repositories for the winning teams and the technical report submitted to the DCASE challenge:</p>\n\n<ul>\n<li>Cochlear.ai - <a href=\"https://github.com/finejuly/dcase2018_task2_cochlearai\">https://github.com/finejuly/dcase2018_task2_cochlearai</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Jeong_102.pdf\">technical report</a></li>\n<li>Matthias Dorfer CPJKU - <a href=\"https://github.com/CPJKU/dcase_task2\">https://github.com/CPJKU/dcase_task2</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Dorfer_999.pdf\">technical report</a></li>\n<li>Surrey CVSSP - <a href=\"https://github.com/turab95/dcase2018_task2\">https://github.com/turab95/dcase2018_task2</a> - <a href=\"http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Iqbal_89.pdf\">technical report</a></li>\n</ul>\n\n<p>We encourage you to visit the <a href=\"http://dcase.community/challenge2018/task-general-purpose-audio-tagging-results\">DCASE Task 2 results page</a> which includes an interesting summary of the characteristics of the systems of the teams who submitted both to Kaggle and the DCASE challenge.</p>\n\n<h3>Judges' award</h3>\n\n<p>After careful inspection of the challenge submissions who met the <a href=\"https://www.kaggle.com/c/freesound-audio-tagging/discussion/59932\">eligibility requirements</a> for the Judges' Award, we chose to award the following submission:</p>\n\n<ul>\n<li><strong>Emilia_NTU</strong> : The report includes an interesting and well-explained approach for dealing with the problem of potentially noisy labels in the non-verified portion of FSDKaggle2018. This self-supervision approach is based on pseudo-labeling and is complemented with label smoothing. Despite model ensembling was utilized (which was not preferred for the Judges' Award), competitive single-models have been considered too as a result of the self-supervision methodology. We’ve valued the insights disclosed from the systematic analysis conducted rather than the scores achieved. Also, the architecture proposed combines the internal representations learned with a CNN, and with few fully connected layers, making the most out of the input log-mel spectrogram. Last but not least we’ve valued the report readability.</li>\n</ul>\n\n<hr>\n\n<p>Aaaaand that's it for the Freesound General-Purpose Audio Tagging Challenge. Congratulations to all participants and specially to the winners! We hope you enjoyed participating as much as we did organizing it. We hope to come back next year with a bigger dataset and an even more challenging competition.</p>\n\n<p>Sincerely,</p>\n\n<p>Eduardo Fonseca, Frederic Font (Freesound, MTG-UPF)</p>\n\n<p>Dan Ellis, Manoj Plakal (Google Machine Perception)</p>",
      "rawMarkdown": "Hi everyone,\n\nThis post is to officially announce the winners of the **Freesound General-Purpose Audio Tagging Challenge** and the Judges' Award. \n\n\n### Challenge winners\n\nThese are the official winners of the competition and their score in the private leaderboard:\n\n1. **Cochlear.ai** - 0.953810\n2. **Matthias Dorfer CPJKU** - 0.951757\n3. **Surrey CVSSP**\t- 0.951244\n\nCongratulations!\n\nOne of the requirements for the winning teams was to open-source the code for the submission. Here is the link to the code repositories for the winning teams and the technical report submitted to the DCASE challenge:\n\n * Cochlear.ai - https://github.com/finejuly/dcase2018_task2_cochlearai - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Jeong_102.pdf)\n * Matthias Dorfer CPJKU - https://github.com/CPJKU/dcase_task2 - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Dorfer_999.pdf)\n * Surrey CVSSP\t- https://github.com/turab95/dcase2018_task2 - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Iqbal_89.pdf)\n\n\n\nWe encourage you to visit the [DCASE Task 2 results page](http://dcase.community/challenge2018/task-general-purpose-audio-tagging-results) which includes an interesting summary of the characteristics of the systems of the teams who submitted both to Kaggle and the DCASE challenge.\n\n\n\n### Judges' award\n\nAfter careful inspection of the challenge submissions who met the [eligibility requirements](https://www.kaggle.com/c/freesound-audio-tagging/discussion/59932) for the Judges' Award, we chose to award the following submission:\n\n * **Emilia_NTU** : The report includes an interesting and well-explained approach for dealing with the problem of potentially noisy labels in the non-verified portion of FSDKaggle2018. This self-supervision approach is based on pseudo-labeling and is complemented with label smoothing. Despite model ensembling was utilized (which was not preferred for the Judges' Award), competitive single-models have been considered too as a result of the self-supervision methodology. We’ve valued the insights disclosed from the systematic analysis conducted rather than the scores achieved. Also, the architecture proposed combines the internal representations learned with a CNN, and with few fully connected layers, making the most out of the input log-mel spectrogram. Last but not least we’ve valued the report readability.\n\n\n\n---\n\nAaaaand that's it for the Freesound General-Purpose Audio Tagging Challenge. Congratulations to all participants and specially to the winners! We hope you enjoyed participating as much as we did organizing it. We hope to come back next year with a bigger dataset and an even more challenging competition.\n\n\nSincerely,\n\nEduardo Fonseca, Frederic Font (Freesound, MTG-UPF)\n\nDan Ellis, Manoj Plakal (Google Machine Perception)\n\n\n\n",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "395467": "Hi everyone,\n\nThis post is to officially announce the winners of the **Freesound General-Purpose Audio Tagging Challenge** and the Judges' Award. \n\n\n### Challenge winners\n\nThese are the official winners of the competition and their score in the private leaderboard:\n\n1. **Cochlear.ai** - 0.953810\n2. **Matthias Dorfer CPJKU** - 0.951757\n3. **Surrey CVSSP**\t- 0.951244\n\nCongratulations!\n\nOne of the requirements for the winning teams was to open-source the code for the submission. Here is the link to the code repositories for the winning teams and the technical report submitted to the DCASE challenge:\n\n * Cochlear.ai - https://github.com/finejuly/dcase2018_task2_cochlearai - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Jeong_102.pdf)\n * Matthias Dorfer CPJKU - https://github.com/CPJKU/dcase_task2 - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Dorfer_999.pdf)\n * Surrey CVSSP\t- https://github.com/turab95/dcase2018_task2 - [technical report](http://dcase.community/documents/challenge2018/technical_reports/DCASE2018_Iqbal_89.pdf)\n\n\n\nWe encourage you to visit the [DCASE Task 2 results page](http://dcase.community/challenge2018/task-general-purpose-audio-tagging-results) which includes an interesting summary of the characteristics of the systems of the teams who submitted both to Kaggle and the DCASE challenge.\n\n\n\n### Judges' award\n\nAfter careful inspection of the challenge submissions who met the [eligibility requirements](https://www.kaggle.com/c/freesound-audio-tagging/discussion/59932) for the Judges' Award, we chose to award the following submission:\n\n * **Emilia_NTU** : The report includes an interesting and well-explained approach for dealing with the problem of potentially noisy labels in the non-verified portion of FSDKaggle2018. This self-supervision approach is based on pseudo-labeling and is complemented with label smoothing. Despite model ensembling was utilized (which was not preferred for the Judges' Award), competitive single-models have been considered too as a result of the self-supervision methodology. We’ve valued the insights disclosed from the systematic analysis conducted rather than the scores achieved. Also, the architecture proposed combines the internal representations learned with a CNN, and with few fully connected layers, making the most out of the input log-mel spectrogram. Last but not least we’ve valued the report readability.\n\n\n\n---\n\nAaaaand that's it for the Freesound General-Purpose Audio Tagging Challenge. Congratulations to all participants and specially to the winners! We hope you enjoyed participating as much as we did organizing it. We hope to come back next year with a bigger dataset and an even more challenging competition.\n\n\nSincerely,\n\nEduardo Fonseca, Frederic Font (Freesound, MTG-UPF)\n\nDan Ellis, Manoj Plakal (Google Machine Perception)\n\n\n\n"
  }
}