{ "id": 506834, "title": "Second place solution", "url": "/competitions/ibiohash-2024-fgvc11/writeups/duong-anh-kiet-second-place-solution", "author_name": "", "post_date": "2024-05-23T13:02:20.163Z", "votes": 1, "comment_count": 1, "views": 0, "content": "

I appreciate the organizers for the engaging competition. To achieve this result, I have tried many methods, which has been a valuable experience for me.
\nSummary:

\n
    \n
  1. First, I developed a model to extract features as a 48-bit hash using the method round(sigmoid(scale*norm(feature))) with a large scale, len(feature)=48, and trained it using ArcFace [1], LMCot [2], and DistanceLayer [3]. However, the results were only around 0.1-0.12. I suspected that 48 bits were too short, so I experimented with increasing it to 128, 512, 1024, etc. The results improved with longer feature lengths, reaching 0.2-0.23 in iBioHash 2023 [4].
    2. \n
  2. We decided to use 1024-dimensional features and apply KMeans to cluster the gallery and query images, generating a random 48-bit hash for each cluster using MD5 [5].
    3. \n
  3. We also tried re-ranking with local features (DELG) [6] after clustering with KMeans. Due to insufficient time, we didn't conduct too many experiments, but with the current ones, there was no improvement in the results.
    4. \n
\n

Code: (I'm working on cleaning up the code)
\nhttps://github.com/ffyyytt/ibio2024
\nI appreciate Google Colab Free TPU for allowing me to run my code.

\n

Reference:
\n[1] Deng, Jiankang, et al. \"Arcface: Additive angular margin loss for deep face recognition.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
\n[2] Duong, Anh-Kiet, Hoang-Lan Nguyen, and Toan-Thinh Truong. \"Large margin cotangent loss for deep similarity learning.\" 2022 International Conference on Advanced Computing and Analytics (ACOMPA). IEEE, 2022.
\n[3] https://www.kaggle.com/code/motono0223/guie-tensorflow-clip-distancelayer-sub-dim64#Distance-Margin-Layer
\n[4] axx, Venti, Wilson. (2023). iBioHash 2023 - FGVC10. Kaggle. https://kaggle.com/competitions/ibiohash-2023-fgvc10
\n[5] https://www.kaggle.com/competitions/ibiohash-2023-fgvc10/discussion/412160
\n[6] Cao, Bingyi, Andre Araujo, and Jack Sim. \"Unifying deep local and global features for image search.\" Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16. Springer International Publishing, 2020.

", "messages": [ { "id": "2830915", "postDate": "05/23/2024 12:39:27", "content": "

I appreciate the organizers for the engaging competition. To achieve this result, I have tried many methods, which has been a valuable experience for me.
\nSummary:

\n
    \n
  1. First, I developed a model to extract features as a 48-bit hash using the method round(sigmoid(scale*norm(feature))) with a large scale, len(feature)=48, and trained it using ArcFace [1], LMCot [2], and DistanceLayer [3]. However, the results were only around 0.1-0.12. I suspected that 48 bits were too short, so I experimented with increasing it to 128, 512, 1024, etc. The results improved with longer feature lengths, reaching 0.2-0.23 in iBioHash 2023 [4].
    2. \n
  2. We decided to use 1024-dimensional features and apply KMeans to cluster the gallery and query images, generating a random 48-bit hash for each cluster using MD5 [5].
    3. \n
  3. We also tried re-ranking with local features (DELG) [6] after clustering with KMeans. Due to insufficient time, we didn't conduct too many experiments, but with the current ones, there was no improvement in the results.
    4. \n
\n

Code: (I'm working on cleaning up the code)
\nhttps://github.com/ffyyytt/ibio2024
\nI appreciate Google Colab Free TPU for allowing me to run my code.

\n

Reference:
\n[1] Deng, Jiankang, et al. \"Arcface: Additive angular margin loss for deep face recognition.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
\n[2] Duong, Anh-Kiet, Hoang-Lan Nguyen, and Toan-Thinh Truong. \"Large margin cotangent loss for deep similarity learning.\" 2022 International Conference on Advanced Computing and Analytics (ACOMPA). IEEE, 2022.
\n[3] https://www.kaggle.com/code/motono0223/guie-tensorflow-clip-distancelayer-sub-dim64#Distance-Margin-Layer
\n[4] axx, Venti, Wilson. (2023). iBioHash 2023 - FGVC10. Kaggle. https://kaggle.com/competitions/ibiohash-2023-fgvc10
\n[5] https://www.kaggle.com/competitions/ibiohash-2023-fgvc10/discussion/412160
\n[6] Cao, Bingyi, Andre Araujo, and Jack Sim. \"Unifying deep local and global features for image search.\" Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16. Springer International Publishing, 2020.

", "rawMarkdown": "I appreciate the organizers for the engaging competition. To achieve this result, I have tried many methods, which has been a valuable experience for me.\n**Summary**:\n1. First, I developed a model to extract features as a 48-bit hash using the method *round*(sigmoid(scale*norm(feature))) with a large scale, len(feature)=48, and trained it using ArcFace [1], LMCot [2], and DistanceLayer [3]. However, the results were only around 0.1-0.12. I suspected that 48 bits were too short, so I experimented with increasing it to 128, 512, 1024, etc. The results improved with longer feature lengths, reaching 0.2-0.23 in iBioHash 2023 [4].\n2. We decided to use 1024-dimensional features and apply KMeans to cluster the gallery and query images, generating a random 48-bit hash for each cluster using MD5 [5].\n3. We also tried re-ranking with local features (DELG) [6] after clustering with KMeans. Due to insufficient time, we didn't conduct too many experiments, but with the current ones, there was no improvement in the results.\n\n**Code**: (I'm working on cleaning up the code)\nhttps://github.com/ffyyytt/ibio2024\nI appreciate Google Colab Free TPU for allowing me to run my code.\n\n**Reference**:\n[1] Deng, Jiankang, et al. \"Arcface: Additive angular margin loss for deep face recognition.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.\n[2] Duong, Anh-Kiet, Hoang-Lan Nguyen, and Toan-Thinh Truong. \"Large margin cotangent loss for deep similarity learning.\" 2022 International Conference on Advanced Computing and Analytics (ACOMPA). IEEE, 2022.\n[3] https://www.kaggle.com/code/motono0223/guie-tensorflow-clip-distancelayer-sub-dim64#Distance-Margin-Layer\n[4] axx, Venti, Wilson. (2023). iBioHash 2023 - FGVC10. Kaggle. https://kaggle.com/competitions/ibiohash-2023-fgvc10\n[5] https://www.kaggle.com/competitions/ibiohash-2023-fgvc10/discussion/412160\n[6] Cao, Bingyi, Andre Araujo, and Jack Sim. \"Unifying deep local and global features for image search.\" Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16. Springer International Publishing, 2020.", "votes": null }, { "id": "2832929", "postDate": "05/24/2024 00:52:06", "content": "

Thanks for the GitHub, the reference \"Large margin cotangent loss for deep similarity learning\" (On the 2022 ACOMPA IEEE. And Congratulations!

", "rawMarkdown": "Thanks for the GitHub, the reference \"Large margin cotangent loss for deep similarity learning\" (On the 2022 ACOMPA IEEE. And Congratulations!", "votes": null } ], "comments": [ { "id": 2832929, "author_name": "mpwolke", "author_url": "", "post_date": "05/24/2024 00:52:06", "content": "

Thanks for the GitHub, the reference \"Large margin cotangent loss for deep similarity learning\" (On the 2022 ACOMPA IEEE. And Congratulations!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2830915": "I appreciate the organizers for the engaging competition. To achieve this result, I have tried many methods, which has been a valuable experience for me.\n**Summary**:\n1. First, I developed a model to extract features as a 48-bit hash using the method *round*(sigmoid(scale*norm(feature))) with a large scale, len(feature)=48, and trained it using ArcFace [1], LMCot [2], and DistanceLayer [3]. However, the results were only around 0.1-0.12. I suspected that 48 bits were too short, so I experimented with increasing it to 128, 512, 1024, etc. The results improved with longer feature lengths, reaching 0.2-0.23 in iBioHash 2023 [4].\n2. We decided to use 1024-dimensional features and apply KMeans to cluster the gallery and query images, generating a random 48-bit hash for each cluster using MD5 [5].\n3. We also tried re-ranking with local features (DELG) [6] after clustering with KMeans. Due to insufficient time, we didn't conduct too many experiments, but with the current ones, there was no improvement in the results.\n\n**Code**: (I'm working on cleaning up the code)\nhttps://github.com/ffyyytt/ibio2024\nI appreciate Google Colab Free TPU for allowing me to run my code.\n\n**Reference**:\n[1] Deng, Jiankang, et al. \"Arcface: Additive angular margin loss for deep face recognition.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.\n[2] Duong, Anh-Kiet, Hoang-Lan Nguyen, and Toan-Thinh Truong. \"Large margin cotangent loss for deep similarity learning.\" 2022 International Conference on Advanced Computing and Analytics (ACOMPA). IEEE, 2022.\n[3] https://www.kaggle.com/code/motono0223/guie-tensorflow-clip-distancelayer-sub-dim64#Distance-Margin-Layer\n[4] axx, Venti, Wilson. (2023). iBioHash 2023 - FGVC10. Kaggle. https://kaggle.com/competitions/ibiohash-2023-fgvc10\n[5] https://www.kaggle.com/competitions/ibiohash-2023-fgvc10/discussion/412160\n[6] Cao, Bingyi, Andre Araujo, and Jack Sim. \"Unifying deep local and global features for image search.\" Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16. Springer International Publishing, 2020.", "2832929": "Thanks for the GitHub, the reference \"Large margin cotangent loss for deep similarity learning\" (On the 2022 ACOMPA IEEE. And Congratulations!" }, "source": "meta" }