Whether you are a competition Grandmaster or a Novice, getting started on a new Kaggle competition takes time and effort. I recently started to understand this competition and it took a while to find the most critical information. The aim of this topic is to help you get started quickly and hopefully save you some time.
\nI assume you have already read the following sections:
\n\nUnderstand the difference between studies and images. Read @dschettler8845 discussion and have a look at his notebook.
\nIn short:
\nThere was a bit of confusion regarding the possibilities to hand-label external data (here the full discussion: Regarding Hand Labeling of Test Set Data).
\nIn short:
\nSome images were duplicates and in some edge cases for the same image there are different annotations. The problem has been fixed for the test-labels but not for the training set. You might have to go through the training data yourself and remove such cases. This discussion from the Competition host might come in handy: Recommendations for handling duplicates on the train dataset.
\nImage classification and object detection is a recurring topic on Kaggle. Read and study the top solutions of past competitions. The following are good starting point:
\nThis is a work in progress. I will update it based on what's going next and on your comments and suggestions. Please, let me know if I'm missing something important or something is wrong. Also, if there are other resources (code or discussion) you think are worth mentioning, please let me know. Happy kaggling to all and thanks for reading!
", "messages": [ { "id": "1368332", "postDate": "06/28/2021 13:49:00", "content": "Whether you are a competition Grandmaster or a Novice, getting started on a new Kaggle competition takes time and effort. I recently started to understand this competition and it took a while to find the most critical information. The aim of this topic is to help you get started quickly and hopefully save you some time.
\nI assume you have already read the following sections:
\n\nUnderstand the difference between studies and images. Read @dschettler8845 discussion and have a look at his notebook.
\nIn short:
\nThere was a bit of confusion regarding the possibilities to hand-label external data (here the full discussion: Regarding Hand Labeling of Test Set Data).
\nIn short:
\nSome images were duplicates and in some edge cases for the same image there are different annotations. The problem has been fixed for the test-labels but not for the training set. You might have to go through the training data yourself and remove such cases. This discussion from the Competition host might come in handy: Recommendations for handling duplicates on the train dataset.
\nImage classification and object detection is a recurring topic on Kaggle. Read and study the top solutions of past competitions. The following are good starting point:
\nThis is a work in progress. I will update it based on what's going next and on your comments and suggestions. Please, let me know if I'm missing something important or something is wrong. Also, if there are other resources (code or discussion) you think are worth mentioning, please let me know. Happy kaggling to all and thanks for reading!
", "rawMarkdown": "Whether you are a competition Grandmaster or a Novice, getting started on a new Kaggle competition takes time and effort. I recently started to understand this competition and it took a while to find the most critical information. The aim of this topic is to help you get started quickly and hopefully save you some time.\n\nI assume you have already read the following sections:\n- [Overview/Description](https://www.kaggle.com/c/siim-covid19-detection/overview)\n- [Overview/Evaluation](https://www.kaggle.com/c/siim-covid19-detection/overview/evaluation)\n- [Data](https://www.kaggle.com/c/siim-covid19-detection/data)\n\n### Studies vs Images\n\nUnderstand the difference between studies and images. Read @dschettler8845 [discussion](https://www.kaggle.com/c/siim-covid19-detection/discussion/240878) and have a look at his [notebook](https://www.kaggle.com/dschettler8845/covid-detection-studies-with-multiple-images-viz). \n\nIn short:\n - studies might have more than one images\n - (see later) for a given study, the annotations might not always coincide\n - For a given **study**, we need to generate a classification (among 4)\n - For a given **image**, we need to localize opacities (through bounding boxes) \n\n### Hand-labeling of external data\n\nThere was a bit of confusion regarding the possibilities to hand-label external data (here the full discussion: [Regarding Hand Labeling of Test Set Data](https://www.kaggle.com/c/siim-covid19-detection/discussion/243804#1341447)). \n\nIn short:\n - [RICORD](https://www.kaggle.com/raddar/ricord-covid19-xray-positive-tests) and [BIMCV](https://bimcv.cipf.es/bimcv-projects/bimcv-covid19/) are public datasets that can be used. Some of the data coming from these two datasets have been used also by the host of this competition. The following dataset can be used as-it-is, but we are not allowed to hand label it.\n\n### Images duplicates and wrong annotations\n\nSome images were duplicates and in some edge cases for the same image there are different annotations. The problem has been fixed for the [test-labels](https://www.kaggle.com/c/siim-covid19-detection/discussion/246296) but not for the training set. You might have to go through the training data yourself and remove such cases. This discussion from the Competition host might come in handy: [Recommendations for handling duplicates on the train dataset](https://www.kaggle.com/c/siim-covid19-detection/discussion/246597).\n\n### Past competitions\n\nImage classification and object detection is a recurring topic on Kaggle. Read and study the top solutions of past competitions. The following are good starting point:\n\n - [Recap top solutions for Xray Medical Imaging comps](https://www.kaggle.com/c/siim-covid19-detection/discussion/239891)\n - [Winning Solutions of Similar Past Kaggle Competitions](https://www.kaggle.com/c/siim-covid19-detection/discussion/239898)\n\n---\n\nThis is a work in progress. I will update it based on what's going next and on your comments and suggestions. Please, let me know if I'm missing something important or something is wrong. Also, if there are other resources (code or discussion) you think are worth mentioning, please let me know. Happy kaggling to all and thanks for reading!", "votes": null }, { "id": "1374374", "postDate": "07/03/2021 08:40:27", "content": "Great Summary.
\nIt's very helpful.
\nThanks for sharing😃
Great Summary.
\nIt's very helpful.
\nThanks for sharing😃