Attaching a paper for relevant image features extraction in display advertising. They have created features like, Brightness, Saturation, Colorfulness, Contrast, Sharpness, Texture, Grayscale simplicity among others.
\n\nIn a kernel shared by @sban, some of these features are there. I think the features mentioned in the paper can be useful.
", "messages": [ { "id": 328059, "postDate": "2018-05-13T08:51:32.163Z", "content": "Attaching a paper for relevant image features extraction in display advertising. They have created features like, Brightness, Saturation, Colorfulness, Contrast, Sharpness, Texture, Grayscale simplicity among others.
\n\nIn a kernel shared by @sban, some of these features are there. I think the features mentioned in the paper can be useful.
", "rawMarkdown": "Attaching a paper for relevant image features extraction in display advertising. They have created features like, Brightness, Saturation, Colorfulness, Contrast, Sharpness, Texture, Grayscale simplicity among others. \n\nIn a kernel shared by [@sban](https://www.kaggle.com/shivamb/ideas-for-image-features-and-image-quality), some of these features are there. I think the features mentioned in the paper can be useful.\n", "votes": 35 }, { "id": 329032, "postDate": "2018-05-15T15:15:47.160Z", "content": "Here are a couple potentially useful kernels from previous competitions that create image features:
\n\nhttps://www.kaggle.com/the1owl/natural-growth-patterns-fractals-of-nature
\n\nhttps://www.kaggle.com/cttsai/ensembling-gbms-lb-203
", "rawMarkdown": "Here are a couple potentially useful kernels from previous competitions that create image features:\n\nhttps://www.kaggle.com/the1owl/natural-growth-patterns-fractals-of-nature\n\nhttps://www.kaggle.com/cttsai/ensembling-gbms-lb-203", "votes": 14, "replies": [ { "id": 332459, "postDate": "2018-05-23T07:21:24.307Z", "content": "I refactored slightly those from the first script and put them in an easy-to-use transformer.\nYou can get them here if you want.
", "rawMarkdown": "I refactored slightly those from the first script and put them in an easy-to-use transformer.\nYou can get them [here][1] if you want.\n\n\n [1]: https://github.com/minerva-ml/open-solution-avito-demand-prediction/blob/master/feature_extraction.py#L605-L686" } ] }, { "id": 328597, "postDate": "2018-05-14T17:18:59.913Z", "content": "Another interesting paper is NIMA: Neural Image Assessment. A technique to rate images based on visual appeal.
\n\nI have already tried using the mean and standard deviation produced by such a model trained on the AVA Dataset as features for my LightGBM model. It did increase my score but I only tried it on a subset of the training data so far so I can't guarantee it helps ;)
", "rawMarkdown": "Another interesting paper is [NIMA: Neural Image Assessment](https://arxiv.org/abs/1709.05424). A technique to rate images based on visual appeal. \n\nI have already tried using the mean and standard deviation produced by such a model trained on the [AVA Dataset](https://computervisiononline.com/dataset/1105138637) as features for my LightGBM model. It did increase my score but I only tried it on a subset of the training data so far so I can't guarantee it helps ;)", "votes": 5, "replies": [ { "id": 333814, "postDate": "2018-05-25T22:04:39.810Z", "content": "Do you have already test this code ? https://github.com/master/nima
", "rawMarkdown": "Do you have already test this code ? https://github.com/master/nima\n" }, { "id": 333832, "postDate": "2018-05-25T22:19:57.327Z", "content": "I used this implementation.
", "rawMarkdown": "I used [this implementation](https://github.com/titu1994/neural-image-assessment)." }, { "id": 333981, "postDate": "2018-05-26T07:54:41.690Z", "content": "Thanks !
", "rawMarkdown": "Thanks !" }, { "id": 334021, "postDate": "2018-05-26T10:11:34.240Z", "content": "thanks. I will give it a try
", "rawMarkdown": "thanks. I will give it a try" }, { "id": 334055, "postDate": "2018-05-26T12:19:23.520Z", "content": "Also takes 24h to get features with GeForce Ti1070...
", "rawMarkdown": "Also takes 24h to get features with GeForce Ti1070... ", "votes": 1 }, { "id": 335700, "postDate": "2018-05-30T07:27:08.687Z", "content": "Did you try it? My first result with NIMA on all images improves my local CV by 0.0002. I've included mean/std for both MobileNet and InceptionResNetV2 as new features. If I check some NIMA scores, it seems to work fine, images with low scores look bad.
", "rawMarkdown": "Did you try it? My first result with NIMA on all images improves my local CV by 0.0002. I've included mean/std for both MobileNet and InceptionResNetV2 as new features. If I check some NIMA scores, it seems to work fine, images with low scores look bad." } ] }, { "id": 328229, "postDate": "2018-05-13T18:26:34.107Z", "content": "Have you used any of those features in your models?
", "rawMarkdown": "Have you used any of those features in your models? ", "votes": 3, "replies": [ { "id": 328866, "postDate": "2018-05-15T07:52:26.017Z", "content": "Not yet
", "rawMarkdown": "Not yet", "votes": 1 }, { "id": 330026, "postDate": "2018-05-17T21:16:29.207Z", "content": "I tried most of the features from the paper and got a very marginal gain (.0001)
", "rawMarkdown": "I tried most of the features from the paper and got a very marginal gain (.0001)", "votes": 8 }, { "id": 334554, "postDate": "2018-05-27T20:23:09.697Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 328289, "postDate": "2018-05-13T21:52:35.097Z", "content": "thanks for sharing. this would get me started.
", "rawMarkdown": "thanks for sharing. this would get me started. ", "votes": 1 }, { "id": 335294, "postDate": "2018-05-29T14:36:06.093Z", "content": "Image features might be extracted by using pretrained models:\nhttps://www.kaggle.com/insaff/vgg-feature-extraction-by-pretrained-model\nhttps://www.kaggle.com/insaff/img-feature-extraction-with-pretrained-resnet
", "rawMarkdown": "Image features might be extracted by using pretrained models:\nhttps://www.kaggle.com/insaff/vgg-feature-extraction-by-pretrained-model\nhttps://www.kaggle.com/insaff/img-feature-extraction-with-pretrained-resnet\n" }, { "id": 333512, "postDate": "2018-05-25T09:54:06.300Z", "rawMarkdown": "", "isDeleted": true }, { "id": 328068, "postDate": "2018-05-13T09:13:38.820Z", "content": "Thanks for sharing Tajendra.
", "rawMarkdown": "Thanks for sharing Tajendra. ", "votes": 1 }, { "id": 392782, "postDate": "2018-09-24T11:01:45.547Z", "content": "Thanks for sharing !
", "rawMarkdown": "Thanks for sharing !" }, { "id": 331750, "postDate": "2018-05-21T20:06:46.900Z", "content": "Thanks a lot, Tajje!!!
", "rawMarkdown": "Thanks a lot, Tajje!!!" } ], "comments": [ { "id": 329032, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-05-15T15:15:47.160000", "content": "Here are a couple potentially useful kernels from previous competitions that create image features:
\n\nhttps://www.kaggle.com/the1owl/natural-growth-patterns-fractals-of-nature
\n\nhttps://www.kaggle.com/cttsai/ensembling-gbms-lb-203
", "votes": 14, "replies": [ { "id": 332459, "author_name": "Jakub Czakon", "author_url": "", "post_date": "2018-05-23T07:21:24.307000", "content": "I refactored slightly those from the first script and put them in an easy-to-use transformer.\nYou can get them here if you want.
", "votes": 0, "replies": [] } ] }, { "id": 328597, "author_name": "Benjamin Minixhofer", "author_url": "", "post_date": "2018-05-14T17:18:59.913000", "content": "Another interesting paper is NIMA: Neural Image Assessment. A technique to rate images based on visual appeal.
\n\nI have already tried using the mean and standard deviation produced by such a model trained on the AVA Dataset as features for my LightGBM model. It did increase my score but I only tried it on a subset of the training data so far so I can't guarantee it helps ;)
", "votes": 5, "replies": [ { "id": 333814, "author_name": "Martin Sam", "author_url": "", "post_date": "2018-05-25T22:04:39.810000", "content": "Do you have already test this code ? https://github.com/master/nima
", "votes": 0, "replies": [] }, { "id": 333832, "author_name": "Benjamin Minixhofer", "author_url": "", "post_date": "2018-05-25T22:19:57.327000", "content": "I used this implementation.
", "votes": 0, "replies": [] }, { "id": 333981, "author_name": "Martin Sam", "author_url": "", "post_date": "2018-05-26T07:54:41.690000", "content": "Thanks !
", "votes": 0, "replies": [] }, { "id": 334021, "author_name": "Dieter", "author_url": "", "post_date": "2018-05-26T10:11:34.240000", "content": "thanks. I will give it a try
", "votes": 0, "replies": [] }, { "id": 334055, "author_name": "Dieter", "author_url": "", "post_date": "2018-05-26T12:19:23.520000", "content": "Also takes 24h to get features with GeForce Ti1070...
", "votes": 1, "replies": [] }, { "id": 335700, "author_name": "MPWARE", "author_url": "", "post_date": "2018-05-30T07:27:08.687000", "content": "Did you try it? My first result with NIMA on all images improves my local CV by 0.0002. I've included mean/std for both MobileNet and InceptionResNetV2 as new features. If I check some NIMA scores, it seems to work fine, images with low scores look bad.
", "votes": 0, "replies": [] } ] }, { "id": 328229, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-05-13T18:26:34.107000", "content": "Have you used any of those features in your models?
", "votes": 3, "replies": [ { "id": 328866, "author_name": "Tajendra Mehta", "author_url": "", "post_date": "2018-05-15T07:52:26.017000", "content": "Not yet
", "votes": 1, "replies": [] }, { "id": 330026, "author_name": "To Train Them Is My Cause", "author_url": "", "post_date": "2018-05-17T21:16:29.207000", "content": "I tried most of the features from the paper and got a very marginal gain (.0001)
", "votes": 8, "replies": [] }, { "id": 334554, "author_name": "", "author_url": "", "post_date": "2018-05-27T20:23:09.697000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 328289, "author_name": "yimacs", "author_url": "", "post_date": "2018-05-13T21:52:35.097000", "content": "thanks for sharing. this would get me started.
", "votes": 1, "replies": [] }, { "id": 335294, "author_name": "Insaf Ashrapov", "author_url": "", "post_date": "2018-05-29T14:36:06.093000", "content": "Image features might be extracted by using pretrained models:\nhttps://www.kaggle.com/insaff/vgg-feature-extraction-by-pretrained-model\nhttps://www.kaggle.com/insaff/img-feature-extraction-with-pretrained-resnet
", "votes": 0, "replies": [] }, { "id": 333512, "author_name": "", "author_url": "", "post_date": "2018-05-25T09:54:06.300000", "content": "", "votes": 0, "replies": [] }, { "id": 328068, "author_name": "Shivam Bansal", "author_url": "", "post_date": "2018-05-13T09:13:38.820000", "content": "Thanks for sharing Tajendra.
", "votes": 1, "replies": [] }, { "id": 392782, "author_name": "Jean-Eudes Peloye", "author_url": "", "post_date": "2018-09-24T11:01:45.547000", "content": "Thanks for sharing !
", "votes": 0, "replies": [] }, { "id": 331750, "author_name": "Burhan ", "author_url": "", "post_date": "2018-05-21T20:06:46.900000", "content": "Thanks a lot, Tajje!!!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "328059": "Attaching a paper for relevant image features extraction in display advertising. They have created features like, Brightness, Saturation, Colorfulness, Contrast, Sharpness, Texture, Grayscale simplicity among others. \n\nIn a kernel shared by [@sban](https://www.kaggle.com/shivamb/ideas-for-image-features-and-image-quality), some of these features are there. I think the features mentioned in the paper can be useful.\n", "329032": "Here are a couple potentially useful kernels from previous competitions that create image features:\n\nhttps://www.kaggle.com/the1owl/natural-growth-patterns-fractals-of-nature\n\nhttps://www.kaggle.com/cttsai/ensembling-gbms-lb-203", "328597": "Another interesting paper is [NIMA: Neural Image Assessment](https://arxiv.org/abs/1709.05424). A technique to rate images based on visual appeal. \n\nI have already tried using the mean and standard deviation produced by such a model trained on the [AVA Dataset](https://computervisiononline.com/dataset/1105138637) as features for my LightGBM model. It did increase my score but I only tried it on a subset of the training data so far so I can't guarantee it helps ;)", "328229": "Have you used any of those features in your models? ", "328289": "thanks for sharing. this would get me started. ", "335294": "Image features might be extracted by using pretrained models:\nhttps://www.kaggle.com/insaff/vgg-feature-extraction-by-pretrained-model\nhttps://www.kaggle.com/insaff/img-feature-extraction-with-pretrained-resnet\n", "333512": "", "328068": "Thanks for sharing Tajendra. ", "392782": "Thanks for sharing !", "331750": "Thanks a lot, Tajje!!!" } }