First, congratulations to the top teams!!! đđđđđđđ What they achieved is incredible!
\nThis year's Landmark Retrieval and Recognition challenges are the first ML competitions that I've ever participated in. Haven't touched anything ML-related for a year, so this was refreshing.
\nUnfortunately, I joined the Retrieval competition late (8 days before the end) and got only around 2 weeks to participate in this one part-time. I've learned a lot during this competition. If there will be a landmark challenge next year, I'll give it a better fight!
\nAlso, I would like to thank @keetar for being an inspiration â what he achieved with just Colab Pro showed me that you don't need fancy hardware to compete, even in such a hard competition with a ton of data. His excellent writeup was the sole reason why I decided to participate in the Recognition track.
\n0.271 and 0.277 models and got silver medal lol lol lol đđđ I also tried some filtering using Places365 VGG model, but it doesn't work out.I realized that I've been spamming the Discussion section a lot, so I'll keep this short:
\nI trained a ResNet152V2 with GeM (p=3) + ArcFace and a EffNetB6 with GAP + ArcFace global descriptors on Cleaned GLDv2. Didn't have time to train to full convergence âšī¸âšī¸âšī¸
For ArcFace, I started with a small margin (i.e. 1e-3), and gradually increased it at the end of every period of Cosine LR, up to 0.31. It was harder for me to train with a fixed margin.
Training hardware: Colab Pro only!
I simply used a publicly available object detector FastRCNN+InceptionResNetV2 trained on Open Images V4 for non-landmark removal. This gave a boost of around +0.005 on pub. LB.
Because I started late (~2 weeks before the end) and didn't had enough time to experiment around, I used Data Echoing technique to speed up training by ~20%.
Standard re-ranking: Retrieval step (using KNN) + Non-Landmark Removal (as described above) + Local Descriptors extraction (using baseline DELG model) + RANSAC.
During inference, I used index embeddings only! I did not have time to generate embeddings for training samples.
What I didn't have time to finish:
\nstop_gradient.My training code is available on Github. I've also published my final submission.
\nReplacing the global descriptor from the Public Recognition Baseline Example with the model from Public Retrieval Baseline will get you 0.4872/0.5081, or a Bronze medal.
Adding a simple non-landmark removal with object detectors will boost the above described baseline-based solution to a Silver medal.
[spoint]: https://arxiv.org/abs/1712.07629
\n[sglue]: https://arxiv.org/abs/1911.11763
\n[fmatches]: https://local-features-tutorial.github.io/pdfs/Local_features_from_paper_to_practice.pdf
First, congratulations to the top teams!!! đđđđđđđ What they achieved is incredible!
\nThis year's Landmark Retrieval and Recognition challenges are the first ML competitions that I've ever participated in. Haven't touched anything ML-related for a year, so this was refreshing.
\nUnfortunately, I joined the Retrieval competition late (8 days before the end) and got only around 2 weeks to participate in this one part-time. I've learned a lot during this competition. If there will be a landmark challenge next year, I'll give it a better fight!
\nAlso, I would like to thank @keetar for being an inspiration â what he achieved with just Colab Pro showed me that you don't need fancy hardware to compete, even in such a hard competition with a ton of data. His excellent writeup was the sole reason why I decided to participate in the Recognition track.
\n0.271 and 0.277 models and got silver medal lol lol lol đđđ I also tried some filtering using Places365 VGG model, but it doesn't work out.I realized that I've been spamming the Discussion section a lot, so I'll keep this short:
\nI trained a ResNet152V2 with GeM (p=3) + ArcFace and a EffNetB6 with GAP + ArcFace global descriptors on Cleaned GLDv2. Didn't have time to train to full convergence âšī¸âšī¸âšī¸
For ArcFace, I started with a small margin (i.e. 1e-3), and gradually increased it at the end of every period of Cosine LR, up to 0.31. It was harder for me to train with a fixed margin.
Training hardware: Colab Pro only!
I simply used a publicly available object detector FastRCNN+InceptionResNetV2 trained on Open Images V4 for non-landmark removal. This gave a boost of around +0.005 on pub. LB.
Because I started late (~2 weeks before the end) and didn't had enough time to experiment around, I used Data Echoing technique to speed up training by ~20%.
Standard re-ranking: Retrieval step (using KNN) + Non-Landmark Removal (as described above) + Local Descriptors extraction (using baseline DELG model) + RANSAC.
During inference, I used index embeddings only! I did not have time to generate embeddings for training samples.
What I didn't have time to finish:
\nstop_gradient.My training code is available on Github. I've also published my final submission.
\nReplacing the global descriptor from the Public Recognition Baseline Example with the model from Public Retrieval Baseline will get you 0.4872/0.5081, or a Bronze medal.
Adding a simple non-landmark removal with object detectors will boost the above described baseline-based solution to a Silver medal.
[spoint]: https://arxiv.org/abs/1712.07629
\n[sglue]: https://arxiv.org/abs/1911.11763
\n[fmatches]: https://local-features-tutorial.github.io/pdfs/Local_features_from_paper_to_practice.pdf
Congrats ! So you used only Global features ?
", "rawMarkdown": "Congrats ! So you used only Global features ?", "votes": null }, { "id": "1032472", "postDate": "09/30/2020 08:30:09", "content": "I was able to train global features. For local features, I didn't have time to train to convergence, so I used publically available local descriptor. Then, I used PyDegensac for feature matching.
", "rawMarkdown": "I was able to train global features. For local features, I didn't have time to train to convergence, so I used publically available local descriptor. Then, I used PyDegensac for feature matching.", "votes": null }, { "id": "1032601", "postDate": "09/30/2020 10:07:11", "content": "Thx for sharing. May I ask what hardware you used for training models?
", "rawMarkdown": "Thx for sharing. May I ask what hardware you used for training models?", "votes": null }, { "id": "1032644", "postDate": "09/30/2020 10:34:52", "content": "Colab Pro TPU v2-8 and private GCS buckets đ I spent more money on milk tea with boba than on training hardware and cloud computing đ
", "rawMarkdown": "Colab Pro TPU v2-8 and private GCS buckets đ I spent more money on milk tea with boba than on training hardware and cloud computing đ", "votes": null }, { "id": "1043540", "postDate": "10/09/2020 04:04:15", "content": "The results of ensembled baselines are interesting!
", "rawMarkdown": "The results of ensembled baselines are interesting!", "votes": null } ], "comments": [ { "id": 1032311, "author_name": "sidneyng", "author_url": "", "post_date": "09/30/2020 05:47:17", "content": "Congrats ! So you used only Global features ?
", "votes": null, "replies": [ { "id": 1032472, "author_name": "chankhavu", "author_url": "", "post_date": "09/30/2020 08:30:09", "content": "I was able to train global features. For local features, I didn't have time to train to convergence, so I used publically available local descriptor. Then, I used PyDegensac for feature matching.
", "votes": null, "replies": [] } ] }, { "id": 1032601, "author_name": "hav4ik", "author_url": "", "post_date": "09/30/2020 10:07:11", "content": "Thx for sharing. May I ask what hardware you used for training models?
", "votes": null, "replies": [ { "id": 1032644, "author_name": "chankhavu", "author_url": "", "post_date": "09/30/2020 10:34:52", "content": "Colab Pro TPU v2-8 and private GCS buckets đ I spent more money on milk tea with boba than on training hardware and cloud computing đ
", "votes": null, "replies": [] } ] }, { "id": 1043540, "author_name": "sjtuwh", "author_url": "", "post_date": "10/09/2020 04:04:15", "content": "The results of ensembled baselines are interesting!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1032283": "First, **congratulations to the top teams!!!** đđđđđđđ What they achieved is incredible! \n\nThis year's [Landmark Retrieval][glret] and [Recognition][glrec] challenges are the first ML competitions that I've ever participated in. Haven't touched anything ML-related for a year, so this was refreshing.\n\nUnfortunately, I joined the Retrieval competition late (8 days before the end) and got only around 2 weeks to participate in this one part-time. I've learned a lot during this competition. If there will be a landmark challenge next year, I'll give it a better fight!\n\nAlso, I would like to thank @keetar for being an inspiration â what he achieved with just Colab Pro showed me that you don't need fancy hardware to compete, even in such a hard competition with a ton of data. His [excellent writeup][keetar-glr] was the sole reason why I decided to participate in the Recognition track.\n\n## Retrieval Track (39th, Silver medal)\n\n- Nothing special, just a weighted ensemble of `0.271` and `0.277` models and got silver medal lol lol lol đđđ I also tried some filtering using Places365 VGG model, but it doesn't work out.\n\n----------------------------------------------------------------\n\n## Recognition Track (22nd, Silver medal) â margin scheduling\n\nI realized that I've been spamming the Discussion section a lot, so I'll keep this short:\n\n- I trained a [ResNet152V2][resnet152v2] with [GeM][gem] (p=3) + [ArcFace][arcface] and a [EffNetB6][effnetb6] with GAP + [ArcFace][arcface] global descriptors on Cleaned GLDv2. Didn't have time to train to full convergence âšī¸âšī¸âšī¸ \n\n- For [ArcFace][arcface], I started with a small margin (i.e. `1e-3`), and gradually increased it at the end of every period of Cosine LR, up to `0.31`. It was harder for me to train with a fixed margin.\n\n- Training hardware: Colab Pro only!\n\n- I simply used a publicly available object detector [FastRCNN+InceptionResNetV2][inceptresnet] trained on [Open Images V4][oims] for non-landmark removal. This gave a boost of around **+0.005** on pub. LB.\n\n- Because I started late (~2 weeks before the end) and didn't had enough time to experiment around, I used [Data Echoing][dataecho] technique to speed up training **by ~20%**.\n\n- Standard re-ranking: Retrieval step (using KNN) + Non-Landmark Removal (as described above) + Local Descriptors extraction (using baseline DELG model) + RANSAC.\n\n- During inference, I used index embeddings only! I did not have time to generate embeddings for training samples.\n\nWhat I didn't have time to finish:\n\n- Local descriptors. I only implemented the local descriptors [36 hours before the end][panictraining].\n- Multi-Gradient Descent ([MGDA-UB][mgda-ub]). I figured this could allow the local descriptor to train alongside with the global ones without `stop_gradient`.\n- Pre-computed embeddings. Ironically, [I was the first who have asked if this is legal][precompdisc].\n\nMy training code is [available on Github](https://github.com/hav4ik/google-landmarks-2020). I've also published my [final submission](https://www.kaggle.com/chankhavu/glob-effb6-e48-rn152-e40-det-frcnn-ss-orig).\n\n----------------------------------------------------------------\n\n## Fun fact: you could've gotten a Silver medal using just baselines!\n\n- Replacing the global descriptor from the [Public Recognition Baseline Example][recbase] with the model from [Public Retrieval Baseline][b277] will get you `0.4872/0.5081`, or a **Bronze** medal.\n\n- Adding a simple non-landmark removal with object detectors will boost the above described baseline-based solution to a **Silver** medal.\n\n\n\n\n[recbase]: https://www.kaggle.com/paulorzp/baseline-landmark-recognition-lb-0-48\n[b277]: https://www.kaggle.com/nvnnghia/main-0806\n[inceptresnet]: https://tfhub.dev/google/faster_rcnn/openimages_v4/inception_resnet_v2/1\n[resnet152v2]: https://www.tensorflow.org/api_docs/python/tf/keras/applications/ResNet152V2\n[effnetb6]: https://github.com/qubvel/efficientnet\n[delg]: https://arxiv.org/pdf/2001.05027.pdf\n[arcface]: https://arxiv.org/abs/1801.07698\n[gem]: https://arxiv.org/abs/1711.02512\n[dataecho]: https://arxiv.org/abs/1907.05550\n[glret]: https://www.kaggle.com/c/landmark-retrieval-2020/\n[glrec]: https://www.kaggle.com/c/landmark-recognition-2020/\n[spoint]: https://arxiv.org/abs/1712.07629\n[sglue]: https://arxiv.org/abs/1911.11763\n[fmatches]: https://local-features-tutorial.github.io/pdfs/Local_features_from_paper_to_practice.pdf\n[cyclegan]: https://junyanz.github.io/CycleGAN/\n[oims]: https://ai.googleblog.com/2020/02/open-images-v6-now-featuring-localized.html\n[panictraining]: https://www.kaggle.com/c/landmark-recognition-2020/discussion/187344\n[mgda-ub]: https://papers.nips.cc/paper/7334-multi-task-learning-as-multi-objective-optimization.pdf\n[precompdisc]: https://www.kaggle.com/c/landmark-recognition-2020/discussion/176697\n[keetar-glr]: https://www.kaggle.com/c/landmark-retrieval-2020/discussion/176037", "1032311": "Congrats ! So you used only Global features ?", "1032472": "I was able to train global features. For local features, I didn't have time to train to convergence, so I used publically available local descriptor. Then, I used PyDegensac for feature matching.", "1032601": "Thx for sharing. May I ask what hardware you used for training models?", "1032644": "Colab Pro TPU v2-8 and private GCS buckets đ I spent more money on milk tea with boba than on training hardware and cloud computing đ", "1043540": "The results of ensembled baselines are interesting!" }, "source": "meta" }