{ "id": 177078, "title": "2nd place solution summary", "url": "/competitions/landmark-retrieval-2020/writeups/bysj-2nd-place-solution-summary", "author_name": "", "post_date": "2021-07-14T02:57:17.573Z", "votes": 35, "comment_count": 16, "views": 0, "content": "

[update]
\nadd google driver download link:
\nhttps://drive.google.com/file/d/1XnzxMOHhzua9tjrAjo-X55ieKVJzRJw_/view?usp=sharing
\n[update]
\nsubmission to arxiv is on hold. We provide detailed information in pdf file, download link: https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf

\n

Method:
\nOur retrieval method for this competition is depicted in Figure 1. We mainly train two models for final submission and each model includes a backbone model for feature extraction and a head model for classification. ResNeSt2692 and Res2Net200 vd are selected as the backbone model since their good performance on ImageNet. Head model includes a pooling layer and two fully connected(fc) layers. The first fc layer is often called embedding layer or whitening layer whose output size is 512. While the output size of second fc layer is corresponding to the class number of training dataset. Instead of using softmax loss for training, we train these models with arcmargin loss. Arcmargin loss is firstly employed in face recognition, we found it works well in retrieval tasks which can produce distinguishing and compact descriptor in landmark.
\n\"\"

\n

The training process mainly consists of three steps. Firstly, we train the two models with resolution 224x224 on GLDv1 dataset which has total 1215498 images of 14950 classes, and GLDv2-clean dataset which has total 1580470 images of 81313 classes. Secondly, these two models are further trained on GLDv2-clean dataset with resolution 448x448, the parameters of arcmargin loss may change during the process. We believe that using large input size is beneficial to extract feature of tiny landmark. However, we have to adopt the training strategy “from small to large” mainly due to the large cost and lack of GPUs. In the final step, some tricks are experimented to increase the performance. We have tried a lot of methods, such as triplet loss finetuning, circle loss finetuning and etc but only “GemPool” and “Fix” strategy are helpful.

\n

Training and test details
\nAt the data level, we first used GLDv1 and GLDv2-clean data with small resolution to train at a large learning rate, and then used GLDv2-clean data with large resolution to train at a small learning rate. The specific details are listed in Table 1. Table 2 shows the results of training with the above strategies. Table 3 lists the mAP@100 score of model ensemble.

\n

\"\"

\n

\"\"

", "messages": [ { "id": "983996", "postDate": "08/24/2020 18:46:12", "content": "

[update]
\nadd google driver download link:
\nhttps://drive.google.com/file/d/1XnzxMOHhzua9tjrAjo-X55ieKVJzRJw_/view?usp=sharing
\n[update]
\nsubmission to arxiv is on hold. We provide detailed information in pdf file, download link: https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf

\n

Method:
\nOur retrieval method for this competition is depicted in Figure 1. We mainly train two models for final submission and each model includes a backbone model for feature extraction and a head model for classification. ResNeSt2692 and Res2Net200 vd are selected as the backbone model since their good performance on ImageNet. Head model includes a pooling layer and two fully connected(fc) layers. The first fc layer is often called embedding layer or whitening layer whose output size is 512. While the output size of second fc layer is corresponding to the class number of training dataset. Instead of using softmax loss for training, we train these models with arcmargin loss. Arcmargin loss is firstly employed in face recognition, we found it works well in retrieval tasks which can produce distinguishing and compact descriptor in landmark.
\n\"\"

\n

The training process mainly consists of three steps. Firstly, we train the two models with resolution 224x224 on GLDv1 dataset which has total 1215498 images of 14950 classes, and GLDv2-clean dataset which has total 1580470 images of 81313 classes. Secondly, these two models are further trained on GLDv2-clean dataset with resolution 448x448, the parameters of arcmargin loss may change during the process. We believe that using large input size is beneficial to extract feature of tiny landmark. However, we have to adopt the training strategy “from small to large” mainly due to the large cost and lack of GPUs. In the final step, some tricks are experimented to increase the performance. We have tried a lot of methods, such as triplet loss finetuning, circle loss finetuning and etc but only “GemPool” and “Fix” strategy are helpful.

\n

Training and test details
\nAt the data level, we first used GLDv1 and GLDv2-clean data with small resolution to train at a large learning rate, and then used GLDv2-clean data with large resolution to train at a small learning rate. The specific details are listed in Table 1. Table 2 shows the results of training with the above strategies. Table 3 lists the mAP@100 score of model ensemble.

\n

\"\"

\n

\"\"

", "rawMarkdown": "**[update]**\nadd google driver download link: \nhttps://drive.google.com/file/d/1XnzxMOHhzua9tjrAjo-X55ieKVJzRJw_/view?usp=sharing\n**[update]**\nsubmission to arxiv is on hold. We provide detailed information in pdf file, download link: https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf\n\n**Method:**\nOur retrieval method for this competition is depicted in Figure 1. We mainly train two models for final submission and each model includes a backbone model for feature extraction and a head model for classification. ResNeSt2692 and Res2Net200 vd are selected as the backbone model since their good performance on ImageNet. Head model includes a pooling layer and two fully connected(fc) layers. The first fc layer is often called embedding layer or whitening layer whose output size is 512. While the output size of second fc layer is corresponding to the class number of training dataset. Instead of using softmax loss for training, we train these models with arcmargin loss. Arcmargin loss is firstly employed in face recognition, we found it works well in retrieval tasks which can produce distinguishing and compact descriptor in landmark.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2F3128c6a35d4df708a26530ff360c3f98%2Ffigure1.png?generation=1598414665139238&alt=media)\n\nThe training process mainly consists of three steps. Firstly, we train the two models with resolution 224x224 on GLDv1 dataset which has total 1215498 images of 14950 classes, and GLDv2-clean dataset which has total 1580470 images of 81313 classes. Secondly, these two models are further trained on GLDv2-clean dataset with resolution 448x448, the parameters of arcmargin loss may change during the process. We believe that using large input size is beneficial to extract feature of tiny landmark. However, we have to adopt the training strategy “from small to large” mainly due to the large cost and lack of GPUs. In the final step, some tricks are experimented to increase the performance. We have tried a lot of methods, such as triplet loss finetuning, circle loss finetuning and etc but only “GemPool” and “Fix” strategy are helpful.\n\n**Training and test details**\nAt the data level, we first used GLDv1 and GLDv2-clean data with small resolution to train at a large learning rate, and then used GLDv2-clean data with large resolution to train at a small learning rate. The specific details are listed in Table 1. Table 2 shows the results of training with the above strategies. Table 3 lists the mAP@100 score of model ensemble.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2F69543b8fa754361b44694b4375eb8a2b%2Ft1.png?generation=1598414755535679&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2Fe70fe15075878b440733828569423c57%2Ft2.png?generation=1598414770976775&alt=media)", "votes": null }, { "id": "984150", "postDate": "08/24/2020 21:55:26", "content": "

Oddsome high-level PDF paper. Congrats all authors.

", "rawMarkdown": "Oddsome high-level PDF paper. Congrats all authors.", "votes": null }, { "id": "984891", "postDate": "08/25/2020 11:03:55", "content": "

Congratulations! :)
\nIf I may ask, how did you deal with different aspect ratios and resolutions at training? Did you just resize them to a square or cropped a central patch and then resized?
\nThanks

", "rawMarkdown": "Congratulations! :)\nIf I may ask, how did you deal with different aspect ratios and resolutions at training? Did you just resize them to a square or cropped a central patch and then resized?\nThanks", "votes": null }, { "id": "985495", "postDate": "08/25/2020 18:57:26", "content": "

Congrats. However, one question from curiosity, why all the profile pic looks unique? Are you all from the same lab? Interestingly found another profile (https://www.kaggle.com/freedomd/competitions), I don't know if this pic holds any meaning though. Thanks.

", "rawMarkdown": "Congrats. However, one question from curiosity, why all the profile pic looks unique? Are you all from the same lab? Interestingly found another profile (https://www.kaggle.com/freedomd/competitions), I don't know if this pic holds any meaning though. Thanks.", "votes": null }, { "id": "985997", "postDate": "08/26/2020 06:26:50", "content": "

Hi all, we have released a personal repo for training features based on pytorch ( https://github.com/feymanpriv/pymetric ) This repo is easy to use for training metric learning and classification features and also provides functions for feature extraction(multi cards) and searching utils on gpus. We aim to build this repo as a base repo for common image retrieval research afterwards.

", "rawMarkdown": "Hi all, we have released a personal repo for training features based on pytorch ( https://github.com/feymanpriv/pymetric ) This repo is easy to use for training metric learning and classification features and also provides functions for feature extraction(multi cards) and searching utils on gpus. We aim to build this repo as a base repo for common image retrieval research afterwards.", "votes": null }, { "id": "986305", "postDate": "08/26/2020 11:09:31", "content": "

thanks~~~~😬

", "rawMarkdown": "thanks~~~~😬", "votes": null }, { "id": "986310", "postDate": "08/26/2020 11:14:12", "content": "

Thanks ! We are not from the same laboratory； profile pic is our idol, of course, she may also be someone else’s idol 😉

", "rawMarkdown": "Thanks ! We are not from the same laboratory； profile pic is our idol, of course, she may also be someone else’s idol 😉", "votes": null }, { "id": "986313", "postDate": "08/26/2020 11:16:50", "content": "

thank you！ We resize image and center crop it to different size as described in the article。https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf

", "rawMarkdown": "thank you！ We resize image and center crop it to different size as described in the article。https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf", "votes": null }, { "id": "986386", "postDate": "08/26/2020 12:45:27", "content": "

Thanks. The paper does not say anything about pre-trained weights. Did you start ResNest from ImageNet or scratch?

", "rawMarkdown": "Thanks. The paper does not say anything about pre-trained weights. Did you start ResNest from ImageNet or scratch?", "votes": null }, { "id": "986392", "postDate": "08/26/2020 12:54:01", "content": "

Yes, we start from pre-trained models(ImageNet).

", "rawMarkdown": "Yes, we start from pre-trained models(ImageNet).", "votes": null }, { "id": "986413", "postDate": "08/26/2020 13:17:37", "content": "

I see, it makes sense.. I was just stunned by the number of epoch you guys used.. 130 epochs is a lot, maybe it's because of the backbone depth

", "rawMarkdown": "I see, it makes sense.. I was just stunned by the number of epoch you guys used.. 130 epochs is a lot, maybe it's because of the backbone depth", "votes": null }, { "id": "987229", "postDate": "08/27/2020 05:21:07", "content": "

Congratulation!!! 么么哒

", "rawMarkdown": "Congratulation!!! 么么哒", "votes": null }, { "id": "987341", "postDate": "08/27/2020 07:08:51", "content": "

🙈

", "rawMarkdown": "🙈", "votes": null }, { "id": "987862", "postDate": "08/27/2020 15:22:56", "content": "

Congratulation!!👍👍👍

", "rawMarkdown": "Congratulation!!👍👍👍", "votes": null }, { "id": "988395", "postDate": "08/28/2020 02:43:28", "content": "

thanks 😜

", "rawMarkdown": "thanks 😜", "votes": null }, { "id": "989955", "postDate": "08/29/2020 08:33:03", "content": "

Thanks. Although I'm using TF, having such a reference repo is definitely helpful!

", "rawMarkdown": "Thanks. Although I'm using TF, having such a reference repo is definitely helpful!", "votes": null }, { "id": "1030296", "postDate": "09/28/2020 14:42:07", "content": "

Your respond is so funny that i can't stop laughing as a Chinese university student.

", "rawMarkdown": "Your respond is so funny that i can't stop laughing as a Chinese university student.", "votes": null } ], "comments": [ { "id": 984150, "author_name": "mpwolke", "author_url": "", "post_date": "08/24/2020 21:55:26", "content": "

Oddsome high-level PDF paper. Congrats all authors.

", "votes": null, "replies": [ { "id": 986305, "author_name": "xuetong1993", "author_url": "", "post_date": "08/26/2020 11:09:31", "content": "

thanks~~~~😬

", "votes": null, "replies": [] } ] }, { "id": 984891, "author_name": "arc144", "author_url": "", "post_date": "08/25/2020 11:03:55", "content": "

Congratulations! :)
\nIf I may ask, how did you deal with different aspect ratios and resolutions at training? Did you just resize them to a square or cropped a central patch and then resized?
\nThanks

", "votes": null, "replies": [ { "id": 986313, "author_name": "xuetong1993", "author_url": "", "post_date": "08/26/2020 11:16:50", "content": "

thank you！ We resize image and center crop it to different size as described in the article。https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf

", "votes": null, "replies": [] }, { "id": 986386, "author_name": "arc144", "author_url": "", "post_date": "08/26/2020 12:45:27", "content": "

Thanks. The paper does not say anything about pre-trained weights. Did you start ResNest from ImageNet or scratch?

", "votes": null, "replies": [] }, { "id": 986392, "author_name": "xuetong1993", "author_url": "", "post_date": "08/26/2020 12:54:01", "content": "

Yes, we start from pre-trained models(ImageNet).

", "votes": null, "replies": [] }, { "id": 986413, "author_name": "arc144", "author_url": "", "post_date": "08/26/2020 13:17:37", "content": "

I see, it makes sense.. I was just stunned by the number of epoch you guys used.. 130 epochs is a lot, maybe it's because of the backbone depth

", "votes": null, "replies": [] } ] }, { "id": 985495, "author_name": "ipythonx", "author_url": "", "post_date": "08/25/2020 18:57:26", "content": "

Congrats. However, one question from curiosity, why all the profile pic looks unique? Are you all from the same lab? Interestingly found another profile (https://www.kaggle.com/freedomd/competitions), I don't know if this pic holds any meaning though. Thanks.

", "votes": null, "replies": [ { "id": 986310, "author_name": "xuetong1993", "author_url": "", "post_date": "08/26/2020 11:14:12", "content": "

Thanks ! We are not from the same laboratory； profile pic is our idol, of course, she may also be someone else’s idol 😉

", "votes": null, "replies": [] }, { "id": 1030296, "author_name": "wcysysu", "author_url": "", "post_date": "09/28/2020 14:42:07", "content": "

Your respond is so funny that i can't stop laughing as a Chinese university student.

", "votes": null, "replies": [] } ] }, { "id": 985997, "author_name": "feymanpriv", "author_url": "", "post_date": "08/26/2020 06:26:50", "content": "

Hi all, we have released a personal repo for training features based on pytorch ( https://github.com/feymanpriv/pymetric ) This repo is easy to use for training metric learning and classification features and also provides functions for feature extraction(multi cards) and searching utils on gpus. We aim to build this repo as a base repo for common image retrieval research afterwards.

", "votes": null, "replies": [ { "id": 989955, "author_name": "chankhavu", "author_url": "", "post_date": "08/29/2020 08:33:03", "content": "

Thanks. Although I'm using TF, having such a reference repo is definitely helpful!

", "votes": null, "replies": [] } ] }, { "id": 987229, "author_name": "fishpam", "author_url": "", "post_date": "08/27/2020 05:21:07", "content": "

Congratulation!!! 么么哒

", "votes": null, "replies": [ { "id": 987341, "author_name": "xuetong1993", "author_url": "", "post_date": "08/27/2020 07:08:51", "content": "

🙈

", "votes": null, "replies": [] } ] }, { "id": 987862, "author_name": "bibeksubedi11", "author_url": "", "post_date": "08/27/2020 15:22:56", "content": "

Congratulation!!👍👍👍

", "votes": null, "replies": [ { "id": 988395, "author_name": "xuetong1993", "author_url": "", "post_date": "08/28/2020 02:43:28", "content": "

thanks 😜

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "983996": "**[update]**\nadd google driver download link: \nhttps://drive.google.com/file/d/1XnzxMOHhzua9tjrAjo-X55ieKVJzRJw_/view?usp=sharing\n**[update]**\nsubmission to arxiv is on hold. We provide detailed information in pdf file, download link: https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf\n\n**Method:**\nOur retrieval method for this competition is depicted in Figure 1. We mainly train two models for final submission and each model includes a backbone model for feature extraction and a head model for classification. ResNeSt2692 and Res2Net200 vd are selected as the backbone model since their good performance on ImageNet. Head model includes a pooling layer and two fully connected(fc) layers. The first fc layer is often called embedding layer or whitening layer whose output size is 512. While the output size of second fc layer is corresponding to the class number of training dataset. Instead of using softmax loss for training, we train these models with arcmargin loss. Arcmargin loss is firstly employed in face recognition, we found it works well in retrieval tasks which can produce distinguishing and compact descriptor in landmark.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2F3128c6a35d4df708a26530ff360c3f98%2Ffigure1.png?generation=1598414665139238&alt=media)\n\nThe training process mainly consists of three steps. Firstly, we train the two models with resolution 224x224 on GLDv1 dataset which has total 1215498 images of 14950 classes, and GLDv2-clean dataset which has total 1580470 images of 81313 classes. Secondly, these two models are further trained on GLDv2-clean dataset with resolution 448x448, the parameters of arcmargin loss may change during the process. We believe that using large input size is beneficial to extract feature of tiny landmark. However, we have to adopt the training strategy “from small to large” mainly due to the large cost and lack of GPUs. In the final step, some tricks are experimented to increase the performance. We have tried a lot of methods, such as triplet loss finetuning, circle loss finetuning and etc but only “GemPool” and “Fix” strategy are helpful.\n\n**Training and test details**\nAt the data level, we first used GLDv1 and GLDv2-clean data with small resolution to train at a large learning rate, and then used GLDv2-clean data with large resolution to train at a small learning rate. The specific details are listed in Table 1. Table 2 shows the results of training with the above strategies. Table 3 lists the mAP@100 score of model ensemble.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2F69543b8fa754361b44694b4375eb8a2b%2Ft1.png?generation=1598414755535679&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3556209%2Fe70fe15075878b440733828569423c57%2Ft2.png?generation=1598414770976775&alt=media)", "984150": "Oddsome high-level PDF paper. Congrats all authors.", "984891": "Congratulations! :)\nIf I may ask, how did you deal with different aspect ratios and resolutions at training? Did you just resize them to a square or cropped a central patch and then resized?\nThanks", "985495": "Congrats. However, one question from curiosity, why all the profile pic looks unique? Are you all from the same lab? Interestingly found another profile (https://www.kaggle.com/freedomd/competitions), I don't know if this pic holds any meaning though. Thanks.", "985997": "Hi all, we have released a personal repo for training features based on pytorch ( https://github.com/feymanpriv/pymetric ) This repo is easy to use for training metric learning and classification features and also provides functions for feature extraction(multi cards) and searching utils on gpus. We aim to build this repo as a base repo for common image retrieval research afterwards.", "986305": "thanks~~~~😬", "986310": "Thanks ! We are not from the same laboratory； profile pic is our idol, of course, she may also be someone else’s idol 😉", "986313": "thank you！ We resize image and center crop it to different size as described in the article。https://vis-bj.bj.bcebos.com/landmark/landmark2020_retrieval.pdf", "986386": "Thanks. The paper does not say anything about pre-trained weights. Did you start ResNest from ImageNet or scratch?", "986392": "Yes, we start from pre-trained models(ImageNet).", "986413": "I see, it makes sense.. I was just stunned by the number of epoch you guys used.. 130 epochs is a lot, maybe it's because of the backbone depth", "987229": "Congratulation!!! 么么哒", "987341": "🙈", "987862": "Congratulation!!👍👍👍", "988395": "thanks 😜", "989955": "Thanks. Although I'm using TF, having such a reference repo is definitely helpful!", "1030296": "Your respond is so funny that i can't stop laughing as a Chinese university student." }, "source": "meta" }