Hi All!
\nFor a long time I have wanted to create summary videos of every competition where I can share interesting bits & tricks from each solution (with due credit, ofcourse).
\nI have been working on the idea for this comp and reading through every solution shared, so I decided to summarise interesting bits in a thread as well:
\nBased on all top readups, I learned YOLO family was a large part of the majority solutions, some teams had also used RCNNs in the ensembles or as single models:
\nCascade RCNN with a backbone: convnext base, I would be super curious to check their code out if they post it I was curious about seeing what Image Augmentations work really well in this setting:
\nI saw many new techniques named, again apologies to everyone who is familiar with these: CLAHE, Image Compression, Coarse Dropout, Mosaic, a few more. Need to learn more about these!
\nMost interestingly, @bestfitting shared they were working on using GANs for image Augmentations! I'm really curious to explore this further and would love to know if anyone has had any success working with this!
\nTricks:
\nThe Devil is in the details and there is a lot of gold too, from inventing a new Augmentation technique, to discovering tricks of reducing GPU usage by splitting images, Decreasing stride to increase model resolution there were many interesting tricks shared:
\nMost imp trick in book of Kaggle: Trusting CV, demonstrated by 1st Pos and everyone that survived the shake up well
\nPS: I have been reading all day in enthusiastic preparation for my chai interviews and summary video, so I will update this once more solutions are shared and getting some sleep 😅
\nThanks everyone and congratulations to everyone for the awesome finish!
", "messages": [ { "id": "1692097", "postDate": "02/15/2022 20:01:35", "content": "Hi All!
\nFor a long time I have wanted to create summary videos of every competition where I can share interesting bits & tricks from each solution (with due credit, ofcourse).
\nI have been working on the idea for this comp and reading through every solution shared, so I decided to summarise interesting bits in a thread as well:
\nBased on all top readups, I learned YOLO family was a large part of the majority solutions, some teams had also used RCNNs in the ensembles or as single models:
\nCascade RCNN with a backbone: convnext base, I would be super curious to check their code out if they post it I was curious about seeing what Image Augmentations work really well in this setting:
\nI saw many new techniques named, again apologies to everyone who is familiar with these: CLAHE, Image Compression, Coarse Dropout, Mosaic, a few more. Need to learn more about these!
\nMost interestingly, @bestfitting shared they were working on using GANs for image Augmentations! I'm really curious to explore this further and would love to know if anyone has had any success working with this!
\nTricks:
\nThe Devil is in the details and there is a lot of gold too, from inventing a new Augmentation technique, to discovering tricks of reducing GPU usage by splitting images, Decreasing stride to increase model resolution there were many interesting tricks shared:
\nMost imp trick in book of Kaggle: Trusting CV, demonstrated by 1st Pos and everyone that survived the shake up well
\nPS: I have been reading all day in enthusiastic preparation for my chai interviews and summary video, so I will update this once more solutions are shared and getting some sleep 😅
\nThanks everyone and congratulations to everyone for the awesome finish!
", "rawMarkdown": "Hi All! \n\nFor a long time I have wanted to create summary videos of every competition where I can share interesting bits & tricks from each solution (with due credit, ofcourse). \n\nI have been working on the idea for this comp and [reading through every solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307598) shared, so I decided to summarise interesting bits in a thread as well:\n\n____\n\n\n### Models:\n\nBased on all top readups, I learned YOLO family was a large part of the majority solutions, some teams had also used RCNNs in the ensembles or as single models:\n\n- Why Yolov5s6 is more effective was discussed [here](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638)\n- [7th Pos solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786) used a *single* model: Custom Cascade RCNN \n- [8th Pos Solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307735) used a `Cascade RCNN with a backbone: convnext base`, I would be super curious to check their code out if they post it \n\n### Image Augmentations\n\nI was curious about seeing what Image Augmentations work really well in this setting: \n\nI saw many new techniques named, again apologies to everyone who is familiar with these: CLAHE, Image Compression, Coarse Dropout, Mosaic, a few more. Need to learn more about these!\n\nMost interestingly, @bestfitting shared they were working on using GANs for image Augmentations! I'm really curious to explore this further and would love to know if anyone has had any success working with this!\n\nTricks:\n\nThe Devil is in the details and there is a lot of gold too, from inventing a new Augmentation technique, to discovering tricks of reducing GPU usage by splitting images, Decreasing stride to increase model resolution there were many interesting tricks shared:\n\n- \"Background Mix Augmentation\" by [3rd Position](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307707): Mixing an image with a BBOX + Image without BBOX to improve generalisation >> Cut paste Aug\n- Reduce GPU Usage: [13th Pos](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307753v)-Upscale images, split them and discard split without bounding box for optimising GPU usage!\n- Progressive Learning\n- Multiscale inference \n- Strong + Weak Augmentation: [7th Place](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786)\n- Decrease stride to increase model resolution: [10th place](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307756) \n\nMost imp trick in book of Kaggle: Trusting CV, demonstrated by 1st Pos and everyone that survived the shake up well\n\nPS: I have been reading all day in enthusiastic preparation for my chai interviews and summary video, so I will update this once more solutions are shared and getting some sleep 😅\n\nThanks everyone and congratulations to everyone for the awesome finish!", "votes": null }, { "id": "1709383", "postDate": "03/02/2022 06:50:30", "content": "Thank you so much for sharing👍
", "rawMarkdown": "Thank you so much for sharing👍", "votes": null }, { "id": "2030090", "postDate": "11/15/2022 07:31:52", "content": "Good work! Thanks for your efforts!
", "rawMarkdown": "Good work! Thanks for your efforts!", "votes": null } ], "comments": [ { "id": 1709383, "author_name": "imnoob", "author_url": "", "post_date": "03/02/2022 06:50:30", "content": "Thank you so much for sharing👍
", "votes": null, "replies": [] }, { "id": 2030090, "author_name": "hridaym25", "author_url": "", "post_date": "11/15/2022 07:31:52", "content": "Good work! Thanks for your efforts!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1692097": "Hi All! \n\nFor a long time I have wanted to create summary videos of every competition where I can share interesting bits & tricks from each solution (with due credit, ofcourse). \n\nI have been working on the idea for this comp and [reading through every solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307598) shared, so I decided to summarise interesting bits in a thread as well:\n\n____\n\n\n### Models:\n\nBased on all top readups, I learned YOLO family was a large part of the majority solutions, some teams had also used RCNNs in the ensembles or as single models:\n\n- Why Yolov5s6 is more effective was discussed [here](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638)\n- [7th Pos solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786) used a *single* model: Custom Cascade RCNN \n- [8th Pos Solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307735) used a `Cascade RCNN with a backbone: convnext base`, I would be super curious to check their code out if they post it \n\n### Image Augmentations\n\nI was curious about seeing what Image Augmentations work really well in this setting: \n\nI saw many new techniques named, again apologies to everyone who is familiar with these: CLAHE, Image Compression, Coarse Dropout, Mosaic, a few more. Need to learn more about these!\n\nMost interestingly, @bestfitting shared they were working on using GANs for image Augmentations! I'm really curious to explore this further and would love to know if anyone has had any success working with this!\n\nTricks:\n\nThe Devil is in the details and there is a lot of gold too, from inventing a new Augmentation technique, to discovering tricks of reducing GPU usage by splitting images, Decreasing stride to increase model resolution there were many interesting tricks shared:\n\n- \"Background Mix Augmentation\" by [3rd Position](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307707): Mixing an image with a BBOX + Image without BBOX to improve generalisation >> Cut paste Aug\n- Reduce GPU Usage: [13th Pos](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307753v)-Upscale images, split them and discard split without bounding box for optimising GPU usage!\n- Progressive Learning\n- Multiscale inference \n- Strong + Weak Augmentation: [7th Place](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307786)\n- Decrease stride to increase model resolution: [10th place](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307756) \n\nMost imp trick in book of Kaggle: Trusting CV, demonstrated by 1st Pos and everyone that survived the shake up well\n\nPS: I have been reading all day in enthusiastic preparation for my chai interviews and summary video, so I will update this once more solutions are shared and getting some sleep 😅\n\nThanks everyone and congratulations to everyone for the awesome finish!", "1709383": "Thank you so much for sharing👍", "2030090": "Good work! Thanks for your efforts!" }, "source": "meta" }