Hello Dear Kagglers!
\nI hope this dataset will be helpful for whatever you want to do about.
\nWe can get know each other. Where are you from? What do you do know? Maybe you can share your Linkedin. This is totally depend to you!
\nFor example, I am Berkay from Turkey and I am working as a data analyst at Getir(\"bring\" in Turkish) that is a Turkish start-up founded in 2015. This is my Linkedin: https://www.linkedin.com/in/berkayalan/
\nYou can also share your project ideas with this data. What do you plan to do? Are you looking for someone to work together?
\nShare your ideas and make your voice heard!💯🔥
", "messages": [ { "id": 1592008, "postDate": "2021-11-22T19:41:55.247Z", "content": "Hello Dear Kagglers!
\nI hope this dataset will be helpful for whatever you want to do about.
\nWe can get know each other. Where are you from? What do you do know? Maybe you can share your Linkedin. This is totally depend to you!
\nFor example, I am Berkay from Turkey and I am working as a data analyst at Getir(\"bring\" in Turkish) that is a Turkish start-up founded in 2015. This is my Linkedin: https://www.linkedin.com/in/berkayalan/
\nYou can also share your project ideas with this data. What do you plan to do? Are you looking for someone to work together?
\nShare your ideas and make your voice heard!💯🔥
", "rawMarkdown": "Hello Dear Kagglers!\n\nI hope this dataset will be helpful for whatever you want to do about.\n\nWe can get know each other. Where are you from? What do you do know? Maybe you can share your Linkedin. This is totally depend to you!\n\nFor example, I am Berkay from Turkey and I am working as a data analyst at Getir(\"bring\" in Turkish) that is a Turkish start-up founded in 2015. This is my Linkedin: https://www.linkedin.com/in/berkayalan/\n\nYou can also share your project ideas with this data. What do you plan to do? Are you looking for someone to work together?\n\nShare your ideas and make your voice heard!💯🔥", "votes": 2 }, { "id": 1592214, "postDate": "2021-11-23T01:25:38.823Z", "content": "Hey all, I’m not very experienced so I plan on doing what I always do: draft @julian3833 and @tensorgirl or team up with an older coder with too many demands and too little time.
", "rawMarkdown": "Hey all, I’m not very experienced so I plan on doing what I always do: draft @julian3833 and @tensorgirl or team up with an older coder with too many demands and too little time. ", "votes": 1 }, { "id": 1594313, "postDate": "2021-11-24T17:40:41.677Z", "content": "I’m Larry. I work in Azure Machine Learning at Microsoft but studied Marine Biology in college and live in Hawai’i, so have a great affinity for the use of ML to aid the marine environment (this competition has actively distracted me from another “real” task that’s also applying ML to the marine environment! :-) ). The more I look at the sample data and think about the problem, the more I want to take advantage of multiple frames, the constant and unidirectional motion of the tow, and the (essentially) non-moving target COTS (objects start near top of frame and move down as they pass under the camera). I also think the distinctive color of the COTS could be an important signal, although the underwater color shifts alter dramatically at distance and are hard to compensate for.
\nI’m considerably more familiar with Pytorch than Tensorflow, although I used and liked Keras in its earlier days. For object detection, I’m most familiar with Detectron2 and like how easy it is to customize for new classes. If I were convinced that single-frame object detection was a respectable strategy, I’d probably (personally):
\nI see the sparsity of the training data as a challenge (no surprise, I suppose).
", "rawMarkdown": "I’m Larry. I work in Azure Machine Learning at Microsoft but studied Marine Biology in college and live in Hawai’i, so have a great affinity for the use of ML to aid the marine environment (this competition has actively distracted me from another “real” task that’s also applying ML to the marine environment! :-) ). The more I look at the sample data and think about the problem, the more I want to take advantage of multiple frames, the constant and unidirectional motion of the tow, and the (essentially) non-moving target COTS (objects start near top of frame and move down as they pass under the camera). I also think the distinctive color of the COTS could be an important signal, although the underwater color shifts alter dramatically at distance and are hard to compensate for.\n\nI’m considerably more familiar with Pytorch than Tensorflow, although I used and liked Keras in its earlier days. For object detection, I’m most familiar with Detectron2 and like how easy it is to customize for new classes. If I were convinced that single-frame object detection was a respectable strategy, I’d probably (personally):\n\n* fine-tune a Detectron2 model to get a result; then\n* maybe use https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN/convert_d2 for conversion; or possibly\n* consider a teacher-student process to get a more efficient model \n\nI see the sparsity of the training data as a challenge (no surprise, I suppose). " }, { "id": 1593620, "postDate": "2021-11-24T05:24:43.500Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1592214, "author_name": "Evan Cofsky", "author_url": "", "post_date": "2021-11-23T01:25:38.823000", "content": "Hey all, I’m not very experienced so I plan on doing what I always do: draft @julian3833 and @tensorgirl or team up with an older coder with too many demands and too little time.
", "votes": 1, "replies": [] }, { "id": 1594313, "author_name": "LarryOBrien", "author_url": "", "post_date": "2021-11-24T17:40:41.677000", "content": "I’m Larry. I work in Azure Machine Learning at Microsoft but studied Marine Biology in college and live in Hawai’i, so have a great affinity for the use of ML to aid the marine environment (this competition has actively distracted me from another “real” task that’s also applying ML to the marine environment! :-) ). The more I look at the sample data and think about the problem, the more I want to take advantage of multiple frames, the constant and unidirectional motion of the tow, and the (essentially) non-moving target COTS (objects start near top of frame and move down as they pass under the camera). I also think the distinctive color of the COTS could be an important signal, although the underwater color shifts alter dramatically at distance and are hard to compensate for.
\nI’m considerably more familiar with Pytorch than Tensorflow, although I used and liked Keras in its earlier days. For object detection, I’m most familiar with Detectron2 and like how easy it is to customize for new classes. If I were convinced that single-frame object detection was a respectable strategy, I’d probably (personally):
\nI see the sparsity of the training data as a challenge (no surprise, I suppose).
", "votes": 0, "replies": [] }, { "id": 1593620, "author_name": "", "author_url": "", "post_date": "2021-11-24T05:24:43.500000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1592008": "Hello Dear Kagglers!\n\nI hope this dataset will be helpful for whatever you want to do about.\n\nWe can get know each other. Where are you from? What do you do know? Maybe you can share your Linkedin. This is totally depend to you!\n\nFor example, I am Berkay from Turkey and I am working as a data analyst at Getir(\"bring\" in Turkish) that is a Turkish start-up founded in 2015. This is my Linkedin: https://www.linkedin.com/in/berkayalan/\n\nYou can also share your project ideas with this data. What do you plan to do? Are you looking for someone to work together?\n\nShare your ideas and make your voice heard!💯🔥", "1592214": "Hey all, I’m not very experienced so I plan on doing what I always do: draft @julian3833 and @tensorgirl or team up with an older coder with too many demands and too little time. ", "1594313": "I’m Larry. I work in Azure Machine Learning at Microsoft but studied Marine Biology in college and live in Hawai’i, so have a great affinity for the use of ML to aid the marine environment (this competition has actively distracted me from another “real” task that’s also applying ML to the marine environment! :-) ). The more I look at the sample data and think about the problem, the more I want to take advantage of multiple frames, the constant and unidirectional motion of the tow, and the (essentially) non-moving target COTS (objects start near top of frame and move down as they pass under the camera). I also think the distinctive color of the COTS could be an important signal, although the underwater color shifts alter dramatically at distance and are hard to compensate for.\n\nI’m considerably more familiar with Pytorch than Tensorflow, although I used and liked Keras in its earlier days. For object detection, I’m most familiar with Detectron2 and like how easy it is to customize for new classes. If I were convinced that single-frame object detection was a respectable strategy, I’d probably (personally):\n\n* fine-tune a Detectron2 model to get a result; then\n* maybe use https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN/convert_d2 for conversion; or possibly\n* consider a teacher-student process to get a more efficient model \n\nI see the sparsity of the training data as a challenge (no surprise, I suppose). ", "1593620": "" } }