So when I first started this competition, I thought a good solution must be able to tackle domain generalization. That is, given the training data from several different sources, we have to generalize (or adapt) to the unseen test data sources. There is a bunch of literature on the topic and I came across a good number of approaches to this problem. Here are some of the articles I looked into:\n - Domain-Adversarial Training of Neural Networks [Link]\n - Learning Transferable Features with Deep Adaptation Networks [Link]\n - Metareg: Towards domain generalization using meta-regularization [Link]\n - Generate To Adapt: Aligning Domains using Generative Adversarial Networks [Link] \n - Domain generalization for object recognition with multi-task autoencoders [Link]\n - Learning from synthetic data: Addressing domain shift for semantic segmentation [Link]
\n\nAfter quite a bit of time wasted on trying to use adversarial networks for this task, I put it on the backburner. I still believe this is an integral part of tackling this competition and I suspect some method could lead to a solid boost in score!
\n\nGood luck!
", "messages": [ { "id": "901041", "postDate": "06/25/2020 08:11:28", "content": "So when I first started this competition, I thought a good solution must be able to tackle domain generalization. That is, given the training data from several different sources, we have to generalize (or adapt) to the unseen test data sources. There is a bunch of literature on the topic and I came across a good number of approaches to this problem. Here are some of the articles I looked into:\n - Domain-Adversarial Training of Neural Networks [Link]\n - Learning Transferable Features with Deep Adaptation Networks [Link]\n - Metareg: Towards domain generalization using meta-regularization [Link]\n - Generate To Adapt: Aligning Domains using Generative Adversarial Networks [Link] \n - Domain generalization for object recognition with multi-task autoencoders [Link]\n - Learning from synthetic data: Addressing domain shift for semantic segmentation [Link]
\n\nAfter quite a bit of time wasted on trying to use adversarial networks for this task, I put it on the backburner. I still believe this is an integral part of tackling this competition and I suspect some method could lead to a solid boost in score!
\n\nGood luck!
", "rawMarkdown": "So when I first started this competition, I thought a good solution must be able to tackle domain generalization. That is, given the training data from several different sources, we have to generalize (or adapt) to the unseen test data sources. There is a bunch of literature on the topic and I came across a good number of approaches to this problem. Here are some of the articles I looked into:\n - Domain-Adversarial Training of Neural Networks [[Link](https://arxiv.org/abs/1505.07818)]\n - Learning Transferable Features with Deep Adaptation Networks [[Link](http://proceedings.mlr.press/v37/long15.pdf)]\n - Metareg: Towards domain generalization using meta-regularization [[Link](https://papers.nips.cc/paper/7378-metareg-towards-domain-generalization-using-meta-regularization.pdf)]\n - Generate To Adapt: Aligning Domains using Generative Adversarial Networks [[Link](http://openaccess.thecvf.com/content_cvpr_2018/papers/Sankaranarayanan_Generate_to_Adapt_CVPR_2018_paper.pdf)] \n - Domain generalization for object recognition with multi-task autoencoders [[Link](http://openaccess.thecvf.com/content_iccv_2015/papers/Ghifary_Domain_Generalization_for_ICCV_2015_paper.pdf)]\n - Learning from synthetic data: Addressing domain shift for semantic segmentation [[Link](http://openaccess.thecvf.com/content_cvpr_2018/papers/Sankaranarayanan_Learning_From_Synthetic_CVPR_2018_paper.pdf)]\n\n\nAfter quite a bit of time wasted on trying to use adversarial networks for this task, I put it on the backburner. I still believe this is an integral part of tackling this competition and I suspect some method could lead to a solid boost in score!\n\nGood luck!", "votes": null }, { "id": "918278", "postDate": "07/07/2020 06:24:58", "content": "thanks for Sharing, I support this idea 👍
", "rawMarkdown": "thanks for Sharing, I support this idea 👍", "votes": null } ], "comments": [ { "id": 918278, "author_name": "tianyu5", "author_url": "", "post_date": "07/07/2020 06:24:58", "content": "thanks for Sharing, I support this idea 👍
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "901041": "So when I first started this competition, I thought a good solution must be able to tackle domain generalization. That is, given the training data from several different sources, we have to generalize (or adapt) to the unseen test data sources. There is a bunch of literature on the topic and I came across a good number of approaches to this problem. Here are some of the articles I looked into:\n - Domain-Adversarial Training of Neural Networks [[Link](https://arxiv.org/abs/1505.07818)]\n - Learning Transferable Features with Deep Adaptation Networks [[Link](http://proceedings.mlr.press/v37/long15.pdf)]\n - Metareg: Towards domain generalization using meta-regularization [[Link](https://papers.nips.cc/paper/7378-metareg-towards-domain-generalization-using-meta-regularization.pdf)]\n - Generate To Adapt: Aligning Domains using Generative Adversarial Networks [[Link](http://openaccess.thecvf.com/content_cvpr_2018/papers/Sankaranarayanan_Generate_to_Adapt_CVPR_2018_paper.pdf)] \n - Domain generalization for object recognition with multi-task autoencoders [[Link](http://openaccess.thecvf.com/content_iccv_2015/papers/Ghifary_Domain_Generalization_for_ICCV_2015_paper.pdf)]\n - Learning from synthetic data: Addressing domain shift for semantic segmentation [[Link](http://openaccess.thecvf.com/content_cvpr_2018/papers/Sankaranarayanan_Learning_From_Synthetic_CVPR_2018_paper.pdf)]\n\n\nAfter quite a bit of time wasted on trying to use adversarial networks for this task, I put it on the backburner. I still believe this is an integral part of tackling this competition and I suspect some method could lead to a solid boost in score!\n\nGood luck!", "918278": "thanks for Sharing, I support this idea 👍" }, "source": "meta" }