We thank all organizers for this very exciting competition.
\nCongratulations to all who finished the competition and to the winners.
We trained CosFace based global features models.
\nThe settings are almost the same as our models in the retrieval2020 (See https://www.kaggle.com/c/landmark-retrieval-2020/discussion/175472)
Additionally, we tried \"Inplace Knowledge Distillation\" by MutualNet paper (https://arxiv.org/pdf/1909.12978.pdf) to train multi-scale image features efficiently.
\nIt has been adopted for all models.
We tried HOW descriptor (https://arxiv.org/pdf/2007.13172.pdf) as an alternative of the global feature model.
\nHOW aggregates CNN based local descriptors into a single global descriptor with ASMK.
\nWe trained a ResNet50-based HOW descriptor model (our pytorch implementation) and got public score = 0.5284 (private = 0.5044).
\nOne of our final submission was based on blending between global cosine similarity and ASMK similarity, that achieved the best public score = 0.6298. However, the other global-feature-only approach achieved the better private score.
| Final submissions | \nPrivate | \nPublic | \n
|---|---|---|
| Global features | \n0.5983 | \n0.6271 | \n
| Global + HOW ASMK | \n0.5960 | \n0.6298 | \n
We thank all organizers for this very exciting competition.
\nCongratulations to all who finished the competition and to the winners.
We trained CosFace based global features models.
\nThe settings are almost the same as our models in the retrieval2020 (See https://www.kaggle.com/c/landmark-retrieval-2020/discussion/175472)
Additionally, we tried \"Inplace Knowledge Distillation\" by MutualNet paper (https://arxiv.org/pdf/1909.12978.pdf) to train multi-scale image features efficiently.
\nIt has been adopted for all models.
We tried HOW descriptor (https://arxiv.org/pdf/2007.13172.pdf) as an alternative of the global feature model.
\nHOW aggregates CNN based local descriptors into a single global descriptor with ASMK.
\nWe trained a ResNet50-based HOW descriptor model (our pytorch implementation) and got public score = 0.5284 (private = 0.5044).
\nOne of our final submission was based on blending between global cosine similarity and ASMK similarity, that achieved the best public score = 0.6298. However, the other global-feature-only approach achieved the better private score.
| Final submissions | \nPrivate | \nPublic | \n
|---|---|---|
| Global features | \n0.5983 | \n0.6271 | \n
| Global + HOW ASMK | \n0.5960 | \n0.6298 | \n
Congratulations with the strong finish! Seems like you guys were also using GLD v2 test set for non-landmark filtering as the 1st place team 😀 I have a question: what is your single model's performance, without filtering (using the organizer's pipeline)?
", "rawMarkdown": "Congratulations with the strong finish! Seems like you guys were also using GLD v2 test set for non-landmark filtering as the 1st place team 😀 I have a question: what is your single model's performance, without filtering (using the organizer's pipeline)?", "votes": null }, { "id": "1034613", "postDate": "10/02/2020 01:02:05", "content": "Thanks.
\nThe differences in performance with and without non-landmark filtering are as follows.
\n| ResNeSt101 single model | \nPrivate | \nPublic | \n
|---|---|---|
| with non-landmark filtering | \n0.5950 | \n0.6148 | \n
| without non-landmark filtering | \n0.5542 | \n0.5777 | \n
(without non-landmark flltering = kNN with private train set only)
\nBut, since our model trained with the addition of a non-landmark class, so this model is likely to have some ability to identify non-landmarks.
", "rawMarkdown": "Thanks.\n\nThe differences in performance with and without non-landmark filtering are as follows.\n\n|ResNeSt101 single model | Private | Public |\n|:------------------------------|:-------|:----------------|\n|with non-landmark filtering |0.5950 |0.6148|\n|without non-landmark filtering|0.5542 |0.5777|\n\n(without non-landmark flltering = kNN with private train set only)\n\nBut, since our model trained with the addition of a non-landmark class, so this model is likely to have some ability to identify non-landmarks.", "votes": null }, { "id": "1034775", "postDate": "10/02/2020 06:58:20", "content": "Thanks a lot for the details!
", "rawMarkdown": "Thanks a lot for the details!", "votes": null }, { "id": "1035006", "postDate": "10/02/2020 11:30:16", "content": "You have done it Great, Would love to know more on it 👍
", "rawMarkdown": "You have done it Great, Would love to know more on it 👍", "votes": null } ], "comments": [ { "id": 1033545, "author_name": "chankhavu", "author_url": "", "post_date": "10/01/2020 04:49:55", "content": "Congratulations with the strong finish! Seems like you guys were also using GLD v2 test set for non-landmark filtering as the 1st place team 😀 I have a question: what is your single model's performance, without filtering (using the organizer's pipeline)?
", "votes": null, "replies": [ { "id": 1034613, "author_name": "knjcode", "author_url": "", "post_date": "10/02/2020 01:02:05", "content": "Thanks.
\nThe differences in performance with and without non-landmark filtering are as follows.
\n| ResNeSt101 single model | \nPrivate | \nPublic | \n
|---|---|---|
| with non-landmark filtering | \n0.5950 | \n0.6148 | \n
| without non-landmark filtering | \n0.5542 | \n0.5777 | \n
(without non-landmark flltering = kNN with private train set only)
\nBut, since our model trained with the addition of a non-landmark class, so this model is likely to have some ability to identify non-landmarks.
", "votes": null, "replies": [] }, { "id": 1034775, "author_name": "chankhavu", "author_url": "", "post_date": "10/02/2020 06:58:20", "content": "Thanks a lot for the details!
", "votes": null, "replies": [] } ] }, { "id": 1035006, "author_name": "ankur95", "author_url": "", "post_date": "10/02/2020 11:30:16", "content": "You have done it Great, Would love to know more on it 👍
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1033538": "We thank all organizers for this very exciting competition. \nCongratulations to all who finished the competition and to the winners.\n\n## Summary\n\n- kNN based on global descriptor\n- Non-landmark filtering with similarity between test set and GLDv2 test set\n- We tried local descriptor methods but they were not used in our final submission\n\n\n## Model Details\n\nWe trained CosFace based global features models. \nThe settings are almost the same as our models in the retrieval2020 (See https://www.kaggle.com/c/landmark-retrieval-2020/discussion/175472)\n\n- Backbones: Ensemble of ResNeSt101, ResNeSt101, ResNeSt200\n- Pooling: GeM (p=3) (Replace GeM p=3 with p=4 in testing)\n- Head: FC->BN->L2\n- Loss: CosFace with Label Smoothing\n- Data Augmentation: HorizontalFlip, RandomResizedCrop, Rotation, RandomGrayScale, ColorJitter, GaussianNoise, Normalize, and GridMask\n- LR: Cosine Annealing LR with warmup, training for 30 epochs + refine 5epochs\n- Input image size in training: 352 (in refine: 640)\n- Adding 600 non-landmark images sampled from the test set of the last year's recognition competition in training (this made the total 81,314 classes).\n\nAdditionally, we tried \"Inplace Knowledge Distillation\" by MutualNet paper (https://arxiv.org/pdf/1909.12978.pdf) to train multi-scale image features efficiently. \nIt has been adopted for all models.\n\n\n## HOW descriptor based ASMK similarity\n\nWe tried HOW descriptor (https://arxiv.org/pdf/2007.13172.pdf) as an alternative of the global feature model. \nHOW aggregates CNN based local descriptors into a single global descriptor with ASMK. \nWe trained a ResNet50-based HOW descriptor model (our pytorch implementation) and got public score = 0.5284 (private = 0.5044). \nOne of our final submission was based on blending between global cosine similarity and ASMK similarity, that achieved the best public score = 0.6298. However, the other global-feature-only approach achieved the better private score.\n\n| Final submissions | Private | Public |\n|-------------------|---------|--------|\n| Global features | 0.5983 | 0.6271 |\n| Global + HOW ASMK | 0.5960 | 0.6298 |\n\n\n## What did not work\n\n- Reranking top-k with HOW ASMK similarity (Blending similarity was better)\n- Reranking top-k with DELG local descriptors\n- DBA\n- Replacing the private train set with GLDv2clean\n - Undersampling up to 350 images per class\n - Removing classes that do not exist in the private train set", "1033545": "Congratulations with the strong finish! Seems like you guys were also using GLD v2 test set for non-landmark filtering as the 1st place team 😀 I have a question: what is your single model's performance, without filtering (using the organizer's pipeline)?", "1034613": "Thanks.\n\nThe differences in performance with and without non-landmark filtering are as follows.\n\n|ResNeSt101 single model | Private | Public |\n|:------------------------------|:-------|:----------------|\n|with non-landmark filtering |0.5950 |0.6148|\n|without non-landmark filtering|0.5542 |0.5777|\n\n(without non-landmark flltering = kNN with private train set only)\n\nBut, since our model trained with the addition of a non-landmark class, so this model is likely to have some ability to identify non-landmarks.", "1034775": "Thanks a lot for the details!", "1035006": "You have done it Great, Would love to know more on it 👍" }, "source": "meta" }