Hello Everyone !!!\nThose who are interested in solving this challenge and just confused where to start...\nI have done a simple research and found a simple path for you...\nStart with easy and effective approach....
\n\ntensorflow\nyolo3: https://github.com/qqwweee/keras-yolo3\nfaster rcnn: object detection in official models of tensorflow models\nMain approach\ndata balancing\ntricks of learning rate\nstudy the rule of scoring\nCode\nhttps://github.com/rabienrose/GoogleAIOpenImg2018\nInclude the necessary files for my project from tensorflow models and keras-yolo3\nExperience\nOne of my friends suddenly asked me to join the challenge one day. I have never involved in any this kind of competition and I have not done any object detection projects before. I only did some classification projects before.Just reading the rule for the competition make me uneasy. But, I like the feeling of pushing myself to the limit, so I started to download the tensorflow models for github.\nI first tried the fast-rcnn model, before it was said the most powerful model current. But it was too slow to train, about 600ms per image with my 1080ti. I only trained with about 20,000 images for couple of epoch. The first submission is very bad, only something like 0.00002. The main result not because the training is bad but because some bugs: I mistake the order of xmin, ymin and so on. After fixed this, I got score like 0.02.\nThe speed of training is a problem, with that speed, I have no time to adjust the model and study the result. So, I decided to move to yolo3. And I also do some data balancing: get rid of images that only contain the very common objects like person, face. Finally, I chose 200,000 images out of 1700,000 train images.\nIt was about 4 hours to train one epoch, I increase the learning rate a little bit if I find the loss enters the platform status. Otherwise I just linearly decrease the learning rate as epoch increase.\nOther configs\nConfidence threshold: 0.01\nExpand the classes based on the hierarchy after prediction, not before training\nNo ensemble (I really did not have time to train multiple models)
", "messages": [ { "id": "542924", "postDate": "06/04/2019 07:51:20", "content": "Hello Everyone !!!\nThose who are interested in solving this challenge and just confused where to start...\nI have done a simple research and found a simple path for you...\nStart with easy and effective approach....
\n\ntensorflow\nyolo3: https://github.com/qqwweee/keras-yolo3\nfaster rcnn: object detection in official models of tensorflow models\nMain approach\ndata balancing\ntricks of learning rate\nstudy the rule of scoring\nCode\nhttps://github.com/rabienrose/GoogleAIOpenImg2018\nInclude the necessary files for my project from tensorflow models and keras-yolo3\nExperience\nOne of my friends suddenly asked me to join the challenge one day. I have never involved in any this kind of competition and I have not done any object detection projects before. I only did some classification projects before.Just reading the rule for the competition make me uneasy. But, I like the feeling of pushing myself to the limit, so I started to download the tensorflow models for github.\nI first tried the fast-rcnn model, before it was said the most powerful model current. But it was too slow to train, about 600ms per image with my 1080ti. I only trained with about 20,000 images for couple of epoch. The first submission is very bad, only something like 0.00002. The main result not because the training is bad but because some bugs: I mistake the order of xmin, ymin and so on. After fixed this, I got score like 0.02.\nThe speed of training is a problem, with that speed, I have no time to adjust the model and study the result. So, I decided to move to yolo3. And I also do some data balancing: get rid of images that only contain the very common objects like person, face. Finally, I chose 200,000 images out of 1700,000 train images.\nIt was about 4 hours to train one epoch, I increase the learning rate a little bit if I find the loss enters the platform status. Otherwise I just linearly decrease the learning rate as epoch increase.\nOther configs\nConfidence threshold: 0.01\nExpand the classes based on the hierarchy after prediction, not before training\nNo ensemble (I really did not have time to train multiple models)
", "rawMarkdown": "Hello Everyone !!!\nThose who are interested in solving this challenge and just confused where to start...\nI have done a simple research and found a simple path for you...\nStart with easy and effective approach....\n\ntensorflow\nyolo3: https://github.com/qqwweee/keras-yolo3\nfaster rcnn: object detection in official models of tensorflow models\nMain approach\ndata balancing\ntricks of learning rate\nstudy the rule of scoring\nCode\nhttps://github.com/rabienrose/GoogleAIOpenImg2018\nInclude the necessary files for my project from tensorflow models and keras-yolo3\nExperience\nOne of my friends suddenly asked me to join the challenge one day. I have never involved in any this kind of competition and I have not done any object detection projects before. I only did some classification projects before.Just reading the rule for the competition make me uneasy. But, I like the feeling of pushing myself to the limit, so I started to download the tensorflow models for github.\nI first tried the fast-rcnn model, before it was said the most powerful model current. But it was too slow to train, about 600ms per image with my 1080ti. I only trained with about 20,000 images for couple of epoch. The first submission is very bad, only something like 0.00002. The main result not because the training is bad but because some bugs: I mistake the order of xmin, ymin and so on. After fixed this, I got score like 0.02.\nThe speed of training is a problem, with that speed, I have no time to adjust the model and study the result. So, I decided to move to yolo3. And I also do some data balancing: get rid of images that only contain the very common objects like person, face. Finally, I chose 200,000 images out of 1700,000 train images.\nIt was about 4 hours to train one epoch, I increase the learning rate a little bit if I find the loss enters the platform status. Otherwise I just linearly decrease the learning rate as epoch increase.\nOther configs\nConfidence threshold: 0.01\nExpand the classes based on the hierarchy after prediction, not before training\nNo ensemble (I really did not have time to train multiple models)", "votes": null }, { "id": "543377", "postDate": "06/04/2019 14:01:20", "content": "You Can follow my kernel to download the dataset from aws server, as it is not directly available on kaggle.\nlink to kernerl:- https://www.kaggle.com/siddhrath/data-fetching-with-aws
", "rawMarkdown": "You Can follow my kernel to download the dataset from aws server, as it is not directly available on kaggle.\nlink to kernerl:- https://www.kaggle.com/siddhrath/data-fetching-with-aws", "votes": null }, { "id": "549870", "postDate": "06/11/2019 05:24:20", "content": "Thanks for sharing!! Really helpful.
", "rawMarkdown": "Thanks for sharing!! Really helpful.", "votes": null } ], "comments": [ { "id": 543377, "author_name": "siddhrath", "author_url": "", "post_date": "06/04/2019 14:01:20", "content": "You Can follow my kernel to download the dataset from aws server, as it is not directly available on kaggle.\nlink to kernerl:- https://www.kaggle.com/siddhrath/data-fetching-with-aws
", "votes": null, "replies": [] }, { "id": 549870, "author_name": "shwetagoyal4", "author_url": "", "post_date": "06/11/2019 05:24:20", "content": "Thanks for sharing!! Really helpful.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "542924": "Hello Everyone !!!\nThose who are interested in solving this challenge and just confused where to start...\nI have done a simple research and found a simple path for you...\nStart with easy and effective approach....\n\ntensorflow\nyolo3: https://github.com/qqwweee/keras-yolo3\nfaster rcnn: object detection in official models of tensorflow models\nMain approach\ndata balancing\ntricks of learning rate\nstudy the rule of scoring\nCode\nhttps://github.com/rabienrose/GoogleAIOpenImg2018\nInclude the necessary files for my project from tensorflow models and keras-yolo3\nExperience\nOne of my friends suddenly asked me to join the challenge one day. I have never involved in any this kind of competition and I have not done any object detection projects before. I only did some classification projects before.Just reading the rule for the competition make me uneasy. But, I like the feeling of pushing myself to the limit, so I started to download the tensorflow models for github.\nI first tried the fast-rcnn model, before it was said the most powerful model current. But it was too slow to train, about 600ms per image with my 1080ti. I only trained with about 20,000 images for couple of epoch. The first submission is very bad, only something like 0.00002. The main result not because the training is bad but because some bugs: I mistake the order of xmin, ymin and so on. After fixed this, I got score like 0.02.\nThe speed of training is a problem, with that speed, I have no time to adjust the model and study the result. So, I decided to move to yolo3. And I also do some data balancing: get rid of images that only contain the very common objects like person, face. Finally, I chose 200,000 images out of 1700,000 train images.\nIt was about 4 hours to train one epoch, I increase the learning rate a little bit if I find the loss enters the platform status. Otherwise I just linearly decrease the learning rate as epoch increase.\nOther configs\nConfidence threshold: 0.01\nExpand the classes based on the hierarchy after prediction, not before training\nNo ensemble (I really did not have time to train multiple models)", "543377": "You Can follow my kernel to download the dataset from aws server, as it is not directly available on kaggle.\nlink to kernerl:- https://www.kaggle.com/siddhrath/data-fetching-with-aws", "549870": "Thanks for sharing!! Really helpful." }, "source": "meta" }