{ "id": 240015, "title": "πŸ’₯πŸ’₯ Key Takeaways of Similar Past Kaggle Competitions πŸ”₯πŸ”₯", "url": "/competitions/siim-covid19-detection/discussion/240015", "author_name": "", "post_date": "2021-05-18T10:07:30.706952500Z", "votes": 60, "comment_count": 8, "views": 0, "content": "

In this post, I summarize some pipelines used by winners of similar past Kaggle competitions. For each competition, I summarize the processing techniques, models, and postprocessing they used. I also provide a list of winner solutions of similar past Kaggle competitions in another post. Please check it πŸ˜ƒ

\n

1. VinBigData Chest X-ray Abnormalities Detection

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note:

\n\n

2. RSNA Pneumonia Detection Challenge

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note

\n\n

3. SIIM-ACR Pneumothorax Segmentation

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note

\n", "messages": [ { "id": "1312905", "postDate": "05/18/2021 10:07:30", "content": "

In this post, I summarize some pipelines used by winners of similar past Kaggle competitions. For each competition, I summarize the processing techniques, models, and postprocessing they used. I also provide a list of winner solutions of similar past Kaggle competitions in another post. Please check it πŸ˜ƒ

\n

1. VinBigData Chest X-ray Abnormalities Detection

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note:

\n\n

2. RSNA Pneumonia Detection Challenge

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note

\n\n

3. SIIM-ACR Pneumothorax Segmentation

\n

Preprocessing Techniques

\n\n

Models

\n\n

Postprocessing Techniques

\n\n

Important Things to Note

\n", "rawMarkdown": "In this post, I summarize some pipelines used by winners of similar past Kaggle competitions. For each competition, I summarize the processing techniques, models, and postprocessing they used. I also provide a list of [winner solutions of similar past Kaggle competitions](https://www.kaggle.com/c/siim-covid19-detection/discussion/239898) in another post. Please check it πŸ˜ƒ\n\n## 1. VinBigData Chest X-ray Abnormalities Detection\n\n**Preprocessing Techniques**\n\n- Annotate part of the NIH data and trained a lung localizer for more efficient usage of GPU memory. The drawback is that diseases outside the lung area are guaranteed missing. [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770)\n- Treat each image-annotation pair as independent data. So both the training and validation data were tripled. [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770)\n\n**Models**\n\n- Detectron2 is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511)\n- YOLOv5 is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511), [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- ResNet101 is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- ResNet152 is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- HourGlass is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- EfficientDet is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511), [5th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229711) , [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n- VFNet is used by [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n**Postprocessing Techniques**\n\n- Ensemble models using [https://github.com/ZFTurbo/Weighted-Boxes-Fusion](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511) [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770), [10th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229629)\n- No Penalty For Adding More Bbox. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n- Calibrate Confidence Scores. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n**Important Things to Note**:\n\n- @morizin analyzes the predictions made for the ensemble both by submitting and also visually inspecting predicted bboxes. [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511)\n- It was very difficult to make reliable Cross-Validation since the test set was generated by different labeling processes and Public LB was not trustworthy due to its small size.\n- @cdeotte use different image size for different models. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n## 2. RSNA Pneumonia Detection Challenge\n\n**Preprocessing Techniques**\n\n- Scale the original images to 512x512 resolution, with 256 resolution @dmytropoplavskiy has seen results degradation and using the full resolution was not as practical with heavier base models. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- Data Augmentation. @dmytropoplavskiy: mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427). @yhirano : x-flip, 90-degree rotation, zoom-in/out and random contrast changes (let k be a random int between [-10, 10], add k over all pixels) [6th place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70650)\n\n**Models**\n\n- [RetinaNet](https://github.com/fizyr/keras-retinanet) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421), [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427), with [Focal Loss](https://arxiv.org/abs/1708.02002) [3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632)\n- [Deformable R-FCN](https://github.com/msracver/Deformable-ConvNets) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- [Deformable Relation Networks](https://github.com/msracver/Relation-Networks-for-Object-Detection) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- Semantic Segmentation using UNet-like network architecture [6th place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70650)\n\n**Postprocessing Techniques**\n\n- Ensemble Boxes using [https://github.com/ahrnbom/ensemble-objdet](https://github.com/ahrnbom/ensemble-objdet). [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- Resize the lengths and widths of the final predictions by 87.5% [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n\n**Important Things to Note**\n\n- For RetinaNet: other backbones. Only the ResNet backbones worked well for us. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- For RetinaNet: pre-training detector heads. It was a lot easier to use ImageNet pre-trained backbones and then just start training the whole network from the beginning. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- For RetinaNet: pre-training the backbone on the pneumonia dataset. No improvement. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- As classification outputs overfit much faster comparing to anchors position/size regression outputs, @dmytropoplavskiy added dropout to anchor and the whole image class outputs. In addition to extra regularization, it helped to achieve the optimal classification and regression results around the same epoch. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- @dmytropoplavskiy added an extra output for smaller anchors (level 2 pyramid layer) to handle smaller boxes and added another classification output predicting the class of the whole image ('No Lung Opacity / Not Normal', 'Normal', 'Lung Opacity'). @dmytropoplavskiy has not used the output but even making the model predict other related functions improved the result. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- @pmcheng aggressively used non-maximum suppression to eliminate any overlapping bounding boxes from each network’s output for a given image. [3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632)\n\n## 3. SIIM-ACR Pneumothorax Segmentation\n\n**Preprocessing Techniques**\n\n- Data Augmentation. [2nd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108009): hflip, scale, rotate, bright, blur. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397): ShiftScaleRotate, RandomBrightnessContrast, ElasticTransform, HorizontalFlip from albumentations.\n- Image Downsize. [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981): 576x576 cropped images. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397): training on random crops with (512, 512) size, inference on (768, 768) size.\n\n**Models**\n\n- Resnet34 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824), [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981), [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397)\n- Resnet50 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824)\n- SeResnext50 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824), [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981)\n- Efficientnet-B3, Efficientnet-b5 are used ny [2nd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108009)\n\n**Postprocessing Techniques**\n\n- `if pred_empty > 0.4 or area(pred_mask) < 800: pred_mask = empty`\nParameters are selected on the validation set. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397)\n\n**Important Things to Note**\n\n- @sneddy use [combo loss](https://github.com/SpaceNetChallenge/SpaceNet_Off_Nadir_Solutions/blob/master/selim_sef/training/losses.py) β†’ combinations of BCE, dice and focal. [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824)", "votes": null }, { "id": "1313661", "postDate": "05/18/2021 16:39:44", "content": "

All the links in this thread are redirecting it to the competition's main discussion page. Is it for me only?

", "rawMarkdown": "All the links in this thread are redirecting it to the competition's main discussion page. Is it for me only?", "votes": null }, { "id": "1313686", "postDate": "05/18/2021 16:52:06", "content": "

Thank you for raising this issue πŸ™. I got the problem because I write on Notion first. I'm on my way to solve it 😁

", "rawMarkdown": "Thank you for raising this issue πŸ™. I got the problem because I write on Notion first. I'm on my way to solve it 😁", "votes": null }, { "id": "1327654", "postDate": "05/29/2021 13:59:15", "content": "

Great summary!

", "rawMarkdown": "Great summary!", "votes": null }, { "id": "1330491", "postDate": "05/31/2021 21:16:09", "content": "

thank you very much. It will be very useful :)

", "rawMarkdown": "thank you very much. It will be very useful :)", "votes": null }, { "id": "1337693", "postDate": "06/05/2021 19:38:51", "content": "

Thank you for sharing πŸ™Œ

", "rawMarkdown": "Thank you for sharing πŸ™Œ", "votes": null }, { "id": "1343072", "postDate": "06/10/2021 02:29:46", "content": "

Thank you, I hope you get a good score tooπŸ‘

", "rawMarkdown": "Thank you, I hope you get a good score tooπŸ‘", "votes": null }, { "id": "1355859", "postDate": "06/18/2021 15:27:36", "content": "

This is great summary, thanks for sharing

", "rawMarkdown": "This is great summary, thanks for sharing", "votes": null }, { "id": "1362030", "postDate": "06/23/2021 07:15:46", "content": "

Thanks for sharing, this is super useful!

", "rawMarkdown": "Thanks for sharing, this is super useful!", "votes": null } ], "comments": [ { "id": 1313661, "author_name": "hassiahk", "author_url": "", "post_date": "05/18/2021 16:39:44", "content": "

All the links in this thread are redirecting it to the competition's main discussion page. Is it for me only?

", "votes": null, "replies": [ { "id": 1313686, "author_name": "mhilmiasyrofi", "author_url": "", "post_date": "05/18/2021 16:52:06", "content": "

Thank you for raising this issue πŸ™. I got the problem because I write on Notion first. I'm on my way to solve it 😁

", "votes": null, "replies": [] } ] }, { "id": 1327654, "author_name": "asmirm", "author_url": "", "post_date": "05/29/2021 13:59:15", "content": "

Great summary!

", "votes": null, "replies": [] }, { "id": 1330491, "author_name": "dpaluszk", "author_url": "", "post_date": "05/31/2021 21:16:09", "content": "

thank you very much. It will be very useful :)

", "votes": null, "replies": [] }, { "id": 1337693, "author_name": "ladcva", "author_url": "", "post_date": "06/05/2021 19:38:51", "content": "

Thank you for sharing πŸ™Œ

", "votes": null, "replies": [] }, { "id": 1343072, "author_name": "shangweichen", "author_url": "", "post_date": "06/10/2021 02:29:46", "content": "

Thank you, I hope you get a good score tooπŸ‘

", "votes": null, "replies": [] }, { "id": 1355859, "author_name": "hawkeat", "author_url": "", "post_date": "06/18/2021 15:27:36", "content": "

This is great summary, thanks for sharing

", "votes": null, "replies": [] }, { "id": 1362030, "author_name": "marti164618", "author_url": "", "post_date": "06/23/2021 07:15:46", "content": "

Thanks for sharing, this is super useful!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1312905": "In this post, I summarize some pipelines used by winners of similar past Kaggle competitions. For each competition, I summarize the processing techniques, models, and postprocessing they used. I also provide a list of [winner solutions of similar past Kaggle competitions](https://www.kaggle.com/c/siim-covid19-detection/discussion/239898) in another post. Please check it πŸ˜ƒ\n\n## 1. VinBigData Chest X-ray Abnormalities Detection\n\n**Preprocessing Techniques**\n\n- Annotate part of the NIH data and trained a lung localizer for more efficient usage of GPU memory. The drawback is that diseases outside the lung area are guaranteed missing. [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770)\n- Treat each image-annotation pair as independent data. So both the training and validation data were tripled. [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770)\n\n**Models**\n\n- Detectron2 is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511)\n- YOLOv5 is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511), [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- ResNet101 is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- ResNet152 is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- HourGlass is used by [2nd place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229696)\n- EfficientDet is used by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511), [5th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229711) , [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n- VFNet is used by [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n**Postprocessing Techniques**\n\n- Ensemble models using [https://github.com/ZFTurbo/Weighted-Boxes-Fusion](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) by [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511) [6th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229770), [10th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229629)\n- No Penalty For Adding More Bbox. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n- Calibrate Confidence Scores. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n**Important Things to Note**:\n\n- @morizin analyzes the predictions made for the ensemble both by submitting and also visually inspecting predicted bboxes. [1st place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/231511)\n- It was very difficult to make reliable Cross-Validation since the test set was generated by different labeling processes and Public LB was not trustworthy due to its small size.\n- @cdeotte use different image size for different models. [7th place](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229637)\n\n## 2. RSNA Pneumonia Detection Challenge\n\n**Preprocessing Techniques**\n\n- Scale the original images to 512x512 resolution, with 256 resolution @dmytropoplavskiy has seen results degradation and using the full resolution was not as practical with heavier base models. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- Data Augmentation. @dmytropoplavskiy: mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427). @yhirano : x-flip, 90-degree rotation, zoom-in/out and random contrast changes (let k be a random int between [-10, 10], add k over all pixels) [6th place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70650)\n\n**Models**\n\n- [RetinaNet](https://github.com/fizyr/keras-retinanet) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421), [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427), with [Focal Loss](https://arxiv.org/abs/1708.02002) [3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632)\n- [Deformable R-FCN](https://github.com/msracver/Deformable-ConvNets) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- [Deformable Relation Networks](https://github.com/msracver/Relation-Networks-for-Object-Detection) is used by [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- Semantic Segmentation using UNet-like network architecture [6th place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70650)\n\n**Postprocessing Techniques**\n\n- Ensemble Boxes using [https://github.com/ahrnbom/ensemble-objdet](https://github.com/ahrnbom/ensemble-objdet). [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- Resize the lengths and widths of the final predictions by 87.5% [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n\n**Important Things to Note**\n\n- For RetinaNet: other backbones. Only the ResNet backbones worked well for us. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- For RetinaNet: pre-training detector heads. It was a lot easier to use ImageNet pre-trained backbones and then just start training the whole network from the beginning. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- For RetinaNet: pre-training the backbone on the pneumonia dataset. No improvement. [1st place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421)\n- As classification outputs overfit much faster comparing to anchors position/size regression outputs, @dmytropoplavskiy added dropout to anchor and the whole image class outputs. In addition to extra regularization, it helped to achieve the optimal classification and regression results around the same epoch. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- @dmytropoplavskiy added an extra output for smaller anchors (level 2 pyramid layer) to handle smaller boxes and added another classification output predicting the class of the whole image ('No Lung Opacity / Not Normal', 'Normal', 'Lung Opacity'). @dmytropoplavskiy has not used the output but even making the model predict other related functions improved the result. [2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427)\n- @pmcheng aggressively used non-maximum suppression to eliminate any overlapping bounding boxes from each network’s output for a given image. [3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632)\n\n## 3. SIIM-ACR Pneumothorax Segmentation\n\n**Preprocessing Techniques**\n\n- Data Augmentation. [2nd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108009): hflip, scale, rotate, bright, blur. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397): ShiftScaleRotate, RandomBrightnessContrast, ElasticTransform, HorizontalFlip from albumentations.\n- Image Downsize. [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981): 576x576 cropped images. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397): training on random crops with (512, 512) size, inference on (768, 768) size.\n\n**Models**\n\n- Resnet34 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824), [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981), [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397)\n- Resnet50 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824)\n- SeResnext50 is used by [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824), [3rd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107981)\n- Efficientnet-B3, Efficientnet-b5 are used ny [2nd place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108009)\n\n**Postprocessing Techniques**\n\n- `if pred_empty > 0.4 or area(pred_mask) < 800: pred_mask = empty`\nParameters are selected on the validation set. [4th place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/108397)\n\n**Important Things to Note**\n\n- @sneddy use [combo loss](https://github.com/SpaceNetChallenge/SpaceNet_Off_Nadir_Solutions/blob/master/selim_sef/training/losses.py) β†’ combinations of BCE, dice and focal. [1st place](https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/107824)", "1313661": "All the links in this thread are redirecting it to the competition's main discussion page. Is it for me only?", "1313686": "Thank you for raising this issue πŸ™. I got the problem because I write on Notion first. I'm on my way to solve it 😁", "1327654": "Great summary!", "1330491": "thank you very much. It will be very useful :)", "1337693": "Thank you for sharing πŸ™Œ", "1343072": "Thank you, I hope you get a good score tooπŸ‘", "1355859": "This is great summary, thanks for sharing", "1362030": "Thanks for sharing, this is super useful!" }, "source": "meta" }