Data augmentation is one of the checklists to deal with the imbalance dataset in this post will cover some of the top winning solutions which contains augmentation technique with model architecture, because the effectiveness of data augmentation techniques can depend on the model architecture, as well as the specific characteristics of the data being used. Some models may be more robust to variations in the input data and may not require as much data augmentation, while others may be more sensitive to these variations and may benefit more from data augmentation. In general, data augmentation can be a useful technique for improving the performance of machine learning models, particularly when working with limited amounts of training data. However, it is important to carefully consider the appropriateness of different data augmentation techniques for a given task and to evaluate their effectiveness on the specific dataset and model being used.
\nVinBigData Chest X-ray Abnormalities Detection
\n2nd place solution by @ivanpan
\nRSNA Pneumonia Detection Challenge
\n2nd place by @dmytropoplavskiy
\nbase model: custom RetinaNet (se-resnext101)
\n512x512 resolution
\naugmentations: Mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes.
\nensemble: NMS
base models: RetinaNet (resnet-50 and resnet-101 ) + focal loss
\n224 x 224 resolution as an abdominal radiologist he considered that high image resolution was not necessary for pneumonia bounding box prediction.
\naugmentations: rotation, translation, scaling, and horizontal flipping + random constants
\nNMS to eliminate any overlapping bounding boxes
Will update this thread going forward!!. That's all for now.
", "messages": [ { "id": 2087187, "postDate": "2023-01-05T11:59:49.597Z", "content": "Data augmentation is one of the checklists to deal with the imbalance dataset in this post will cover some of the top winning solutions which contains augmentation technique with model architecture, because the effectiveness of data augmentation techniques can depend on the model architecture, as well as the specific characteristics of the data being used. Some models may be more robust to variations in the input data and may not require as much data augmentation, while others may be more sensitive to these variations and may benefit more from data augmentation. In general, data augmentation can be a useful technique for improving the performance of machine learning models, particularly when working with limited amounts of training data. However, it is important to carefully consider the appropriateness of different data augmentation techniques for a given task and to evaluate their effectiveness on the specific dataset and model being used.
\nVinBigData Chest X-ray Abnormalities Detection
\n2nd place solution by @ivanpan
\nRSNA Pneumonia Detection Challenge
\n2nd place by @dmytropoplavskiy
\nbase model: custom RetinaNet (se-resnext101)
\n512x512 resolution
\naugmentations: Mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes.
\nensemble: NMS
base models: RetinaNet (resnet-50 and resnet-101 ) + focal loss
\n224 x 224 resolution as an abdominal radiologist he considered that high image resolution was not necessary for pneumonia bounding box prediction.
\naugmentations: rotation, translation, scaling, and horizontal flipping + random constants
\nNMS to eliminate any overlapping bounding boxes
Will update this thread going forward!!. That's all for now.
", "rawMarkdown": "Data augmentation is one of the checklists to deal with the imbalance dataset in this post will cover some of the top winning solutions which contains augmentation technique with model architecture, because the effectiveness of data augmentation techniques can depend on the model architecture, as well as the specific characteristics of the data being used. Some models may be more robust to variations in the input data and may not require as much data augmentation, while others may be more sensitive to these variations and may benefit more from data augmentation. In general, data augmentation can be a useful technique for improving the performance of machine learning models, particularly when working with limited amounts of training data. However, it is important to carefully consider the appropriateness of different data augmentation techniques for a given task and to evaluate their effectiveness on the specific dataset and model being used.\n\n**[VinBigData Chest X-ray Abnormalities Detection](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/overview)**\n\n[2nd place solution](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229740) by @ivanpan\n- Based on MMDetection: Cascade_RFP_R50 , GFL_R101 , GFL_X101 , RetinaNet_X101.\n- Base size 1024x1024 , training with FP16.\n- Albumentations like: ShiftScaleRotate, IAAAffine, Blur/GaussianBlur/MedianBlur, RandomBrightnessContrast, IAAAdditiveGaussianNoise/GaussNoise, HorizontalFlip.\n\n**[RSNA Pneumonia Detection Challenge](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/)**\n\n[2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427) by @dmytropoplavskiy\n\nbase model: custom RetinaNet (se-resnext101)\n512x512 resolution\naugmentations: Mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes.\nensemble: NMS\n\n\n[3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632) + [code](https://github.com/pmcheng/rsna-pneumonia) by @pmcheng\n\nbase models: RetinaNet (resnet-50 and resnet-101 ) + focal loss\n224 x 224 resolution as an abdominal radiologist he considered that high image resolution was not necessary for pneumonia bounding box prediction.\naugmentations: rotation, translation, scaling, and horizontal flipping + random constants\nNMS to eliminate any overlapping bounding boxes\n\n\nWill update this thread going forward!!. That's all for now.\n\n\n\n", "votes": 15 }, { "id": 2088580, "postDate": "2023-01-06T12:56:40.707Z", "content": "Thanks a lot for sharing this 💯
", "rawMarkdown": "Thanks a lot for sharing this 💯" } ], "comments": [ { "id": 2088580, "author_name": "Adityam Ghosh", "author_url": "", "post_date": "2023-01-06T12:56:40.707000", "content": "Thanks a lot for sharing this 💯
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2087187": "Data augmentation is one of the checklists to deal with the imbalance dataset in this post will cover some of the top winning solutions which contains augmentation technique with model architecture, because the effectiveness of data augmentation techniques can depend on the model architecture, as well as the specific characteristics of the data being used. Some models may be more robust to variations in the input data and may not require as much data augmentation, while others may be more sensitive to these variations and may benefit more from data augmentation. In general, data augmentation can be a useful technique for improving the performance of machine learning models, particularly when working with limited amounts of training data. However, it is important to carefully consider the appropriateness of different data augmentation techniques for a given task and to evaluate their effectiveness on the specific dataset and model being used.\n\n**[VinBigData Chest X-ray Abnormalities Detection](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/overview)**\n\n[2nd place solution](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/229740) by @ivanpan\n- Based on MMDetection: Cascade_RFP_R50 , GFL_R101 , GFL_X101 , RetinaNet_X101.\n- Base size 1024x1024 , training with FP16.\n- Albumentations like: ShiftScaleRotate, IAAAffine, Blur/GaussianBlur/MedianBlur, RandomBrightnessContrast, IAAAdditiveGaussianNoise/GaussNoise, HorizontalFlip.\n\n**[RSNA Pneumonia Detection Challenge](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/)**\n\n[2nd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70427) by @dmytropoplavskiy\n\nbase model: custom RetinaNet (se-resnext101)\n512x512 resolution\naugmentations: Mild rotations (up to 6 deg), shift, scale, shear and h_flip, for some images random level of blur and noise and gamma changes.\nensemble: NMS\n\n\n[3rd place](https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70632) + [code](https://github.com/pmcheng/rsna-pneumonia) by @pmcheng\n\nbase models: RetinaNet (resnet-50 and resnet-101 ) + focal loss\n224 x 224 resolution as an abdominal radiologist he considered that high image resolution was not necessary for pneumonia bounding box prediction.\naugmentations: rotation, translation, scaling, and horizontal flipping + random constants\nNMS to eliminate any overlapping bounding boxes\n\n\nWill update this thread going forward!!. That's all for now.\n\n\n\n", "2088580": "Thanks a lot for sharing this 💯" } }