In this competition, it is not only important to predict the landmark for an image but also to make sure that if an image is not a landmark that we don't make a prediction.
\n\nThis technique is a necessary ingredient to achieve a better score on LB and for better models. Skimming through previous competitions solution's I found that there are many good and sophisticated solutions to this problem. One of the many solutions is using the Places365 dataset, in its extended version of the dataset contains indoor/outdoor labels which can act as landmark/non-landmark.
\n\nThen I found this notebook from Robin Smith - Keras Landmark or Non-Landmark Identification and I extended the idea to implement the same in PyTorch but found most of the code available at https://github.com/CSAILVision/places365 and pre-trained models as well.
\n\nI just compiled everything for this competition data and the results were amazing.\nThis just contains the Landmark/Non-Landmark classes for test-data but soon will be creating end-to-end training/inference notebook using these landmark/non landmark label.
\n\nYou can read my notebook here PyTorch Landmark or Non-Landmark identification.
\n\nThanks
", "messages": [ { "id": 955056, "postDate": "2020-08-02T09:35:23.100Z", "content": "In this competition, it is not only important to predict the landmark for an image but also to make sure that if an image is not a landmark that we don't make a prediction.
\n\nThis technique is a necessary ingredient to achieve a better score on LB and for better models. Skimming through previous competitions solution's I found that there are many good and sophisticated solutions to this problem. One of the many solutions is using the Places365 dataset, in its extended version of the dataset contains indoor/outdoor labels which can act as landmark/non-landmark.
\n\nThen I found this notebook from Robin Smith - Keras Landmark or Non-Landmark Identification and I extended the idea to implement the same in PyTorch but found most of the code available at https://github.com/CSAILVision/places365 and pre-trained models as well.
\n\nI just compiled everything for this competition data and the results were amazing.\nThis just contains the Landmark/Non-Landmark classes for test-data but soon will be creating end-to-end training/inference notebook using these landmark/non landmark label.
\n\nYou can read my notebook here PyTorch Landmark or Non-Landmark identification.
\n\nThanks
", "rawMarkdown": "In this competition, it is not only important to predict the landmark for an image but also to make sure that if an image is not a landmark that we don't make a prediction.\n\n1. Correctly identifying non-landmark images in the test set will increase the score of a submission.\n2. Removing as much of the non-landmark images from the training set will decrease the total amount of images that we need to train on and it will improve the 'correctness' of the model if it is trained on the landmark images and not on all the selfies, hotel-rooms and beds etcetera that was made near the landmark.\n\nThis technique is a necessary ingredient to achieve a better score on LB and for better models. Skimming through previous competitions solution's I found that there are many good and sophisticated solutions to this problem. One of the many solutions is using the Places365 dataset, in its extended version of the dataset contains indoor/outdoor labels which can act as landmark/non-landmark. \n\nThen I found this notebook from [Robin Smith - Keras Landmark or Non-Landmark Identification](https://www.kaggle.com/rsmits/keras-landmark-or-non-landmark-identification) and I extended the idea to implement the same in PyTorch but found most of the code available at https://github.com/CSAILVision/places365 and pre-trained models as well. \n\nI just compiled everything for this competition data and the results were amazing.\nThis just contains the Landmark/Non-Landmark classes for test-data but soon will be creating end-to-end training/inference notebook using these landmark/non landmark label.\n\n**You can read my notebook here** [PyTorch Landmark or Non-Landmark identification](https://www.kaggle.com/rhtsingh/pytorch-landmark-or-non-landmark-identification).\n\nThanks", "votes": 5 } ], "comments": [], "raw_markdown_by_id": { "955056": "In this competition, it is not only important to predict the landmark for an image but also to make sure that if an image is not a landmark that we don't make a prediction.\n\n1. Correctly identifying non-landmark images in the test set will increase the score of a submission.\n2. Removing as much of the non-landmark images from the training set will decrease the total amount of images that we need to train on and it will improve the 'correctness' of the model if it is trained on the landmark images and not on all the selfies, hotel-rooms and beds etcetera that was made near the landmark.\n\nThis technique is a necessary ingredient to achieve a better score on LB and for better models. Skimming through previous competitions solution's I found that there are many good and sophisticated solutions to this problem. One of the many solutions is using the Places365 dataset, in its extended version of the dataset contains indoor/outdoor labels which can act as landmark/non-landmark. \n\nThen I found this notebook from [Robin Smith - Keras Landmark or Non-Landmark Identification](https://www.kaggle.com/rsmits/keras-landmark-or-non-landmark-identification) and I extended the idea to implement the same in PyTorch but found most of the code available at https://github.com/CSAILVision/places365 and pre-trained models as well. \n\nI just compiled everything for this competition data and the results were amazing.\nThis just contains the Landmark/Non-Landmark classes for test-data but soon will be creating end-to-end training/inference notebook using these landmark/non landmark label.\n\n**You can read my notebook here** [PyTorch Landmark or Non-Landmark identification](https://www.kaggle.com/rhtsingh/pytorch-landmark-or-non-landmark-identification).\n\nThanks" } }