The dataset comes with a rasterized version of a semantic map, which is a 40k by 40k RGB image.
\n\nA week ago I uploaded the reference model, as an example of how to use this semantic map image there is now also a version that uses the map as additional input. In my (limited) experiments I saw an improvement to both precision and recall, especially at larger distances from the ego vehicle.
\n\nUnfortunately the current reference model implementation using the semantic map runs out of memory in a GPU kernel (the dataset and model itself already uses a lot of its memory). Instead, you can find the kernel with output in the SDK repository: https://github.com/lyft/nuscenes-devkit/blob/master/notebooks/Reference%20Model.ipynb
\n\nI hope it helps some of you, happy to answer questions here :)
", "messages": [ { "id": "633082", "postDate": "09/24/2019 12:22:56", "content": "The dataset comes with a rasterized version of a semantic map, which is a 40k by 40k RGB image.
\n\nA week ago I uploaded the reference model, as an example of how to use this semantic map image there is now also a version that uses the map as additional input. In my (limited) experiments I saw an improvement to both precision and recall, especially at larger distances from the ego vehicle.
\n\nUnfortunately the current reference model implementation using the semantic map runs out of memory in a GPU kernel (the dataset and model itself already uses a lot of its memory). Instead, you can find the kernel with output in the SDK repository: https://github.com/lyft/nuscenes-devkit/blob/master/notebooks/Reference%20Model.ipynb
\n\nI hope it helps some of you, happy to answer questions here :)
", "rawMarkdown": "The dataset comes with a rasterized version of a semantic map, which is a 40k by 40k RGB image.\n\nA week ago I uploaded the reference model, as an example of how to use this semantic map image there is now also a version that uses the map as additional input. In my (limited) experiments I saw an improvement to both precision and recall, especially at larger distances from the ego vehicle.\n\nUnfortunately the current reference model implementation using the semantic map runs out of memory in a GPU kernel (the dataset and model itself already uses a lot of its memory). Instead, you can find the kernel with output in the SDK repository: https://github.com/lyft/nuscenes-devkit/blob/master/notebooks/Reference%20Model.ipynb \n\nI hope it helps some of you, happy to answer questions here :)", "votes": null }, { "id": "633119", "postDate": "09/24/2019 13:02:04", "content": "@gzuidhof thank you so much! Do you want people to contribute to the github? I mean, I'd like to finish my best EDA and do a pull request :p
", "rawMarkdown": "gzuidhof thank you so much! Do you want people to contribute to the github? I mean, I'd like to finish my best EDA and do a pull request :p", "votes": null }, { "id": "633145", "postDate": "09/24/2019 13:38:07", "content": "We are very happy to accept contributions! Vladimir is coordinating the SDK and contributions to it.
\n\nAlthough, in my own opinion, Kaggle kernels is probably the best place for EDAs - there is a lot more visibility here, they show up in Google results, and it's easy for people to comment, fork and improve.
", "rawMarkdown": "We are very happy to accept contributions! Vladimir is coordinating the SDK and contributions to it.\n\nAlthough, in my own opinion, Kaggle kernels is probably the best place for EDAs - there is a lot more visibility here, they show up in Google results, and it's easy for people to comment, fork and improve.", "votes": null }, { "id": "636970", "postDate": "09/30/2019 12:54:56", "content": "looks great!
", "rawMarkdown": "looks great!", "votes": null }, { "id": "662038", "postDate": "10/31/2019 02:45:51", "content": "@gzuidhof Could you let us know what are the machine specs used for training the model using the semantic map
", "rawMarkdown": "gzuidhof Could you let us know what are the machine specs used for training the model using the semantic map", "votes": null }, { "id": "664031", "postDate": "11/03/2019 04:29:00", "content": "@gzuidhof - Thanks for sharing the code. When I am trying to execute the code, I am getting index tuple out of range error when get_semantic_map_around_ego function is being called. Is there anything I am missing
", "rawMarkdown": "gzuidhof - Thanks for sharing the code. When I am trying to execute the code, I am getting index tuple out of range error when get_semantic_map_around_ego function is being called. Is there anything I am missing", "votes": null } ], "comments": [ { "id": 633119, "author_name": "jesucristo", "author_url": "", "post_date": "09/24/2019 13:02:04", "content": "@gzuidhof thank you so much! Do you want people to contribute to the github? I mean, I'd like to finish my best EDA and do a pull request :p
", "votes": null, "replies": [ { "id": 633145, "author_name": "gzuidhof", "author_url": "", "post_date": "09/24/2019 13:38:07", "content": "We are very happy to accept contributions! Vladimir is coordinating the SDK and contributions to it.
\n\nAlthough, in my own opinion, Kaggle kernels is probably the best place for EDAs - there is a lot more visibility here, they show up in Google results, and it's easy for people to comment, fork and improve.
", "votes": null, "replies": [] } ] }, { "id": 636970, "author_name": "krishna5555", "author_url": "", "post_date": "09/30/2019 12:54:56", "content": "looks great!
", "votes": null, "replies": [] }, { "id": 662038, "author_name": "venkat555", "author_url": "", "post_date": "10/31/2019 02:45:51", "content": "@gzuidhof Could you let us know what are the machine specs used for training the model using the semantic map
", "votes": null, "replies": [] }, { "id": 664031, "author_name": "manojprabhaakr", "author_url": "", "post_date": "11/03/2019 04:29:00", "content": "@gzuidhof - Thanks for sharing the code. When I am trying to execute the code, I am getting index tuple out of range error when get_semantic_map_around_ego function is being called. Is there anything I am missing
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "633082": "The dataset comes with a rasterized version of a semantic map, which is a 40k by 40k RGB image.\n\nA week ago I uploaded the reference model, as an example of how to use this semantic map image there is now also a version that uses the map as additional input. In my (limited) experiments I saw an improvement to both precision and recall, especially at larger distances from the ego vehicle.\n\nUnfortunately the current reference model implementation using the semantic map runs out of memory in a GPU kernel (the dataset and model itself already uses a lot of its memory). Instead, you can find the kernel with output in the SDK repository: https://github.com/lyft/nuscenes-devkit/blob/master/notebooks/Reference%20Model.ipynb \n\nI hope it helps some of you, happy to answer questions here :)", "633119": "gzuidhof thank you so much! Do you want people to contribute to the github? I mean, I'd like to finish my best EDA and do a pull request :p", "633145": "We are very happy to accept contributions! Vladimir is coordinating the SDK and contributions to it.\n\nAlthough, in my own opinion, Kaggle kernels is probably the best place for EDAs - there is a lot more visibility here, they show up in Google results, and it's easy for people to comment, fork and improve.", "636970": "looks great!", "662038": "gzuidhof Could you let us know what are the machine specs used for training the model using the semantic map", "664031": "gzuidhof - Thanks for sharing the code. When I am trying to execute the code, I am getting index tuple out of range error when get_semantic_map_around_ego function is being called. Is there anything I am missing" }, "source": "meta" }