![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3617078%2F3747eec8c06565f26d428ca2be12edb7%2F2020-01-02%209.12.15.png?generation=1577970922249465&alt=media\")
I was trying to find the relationship between id in train.csv and the car model label in Apollo dataset api repo as following link:\nhttps://github.com/ApolloScapeAuto/dataset-api/blob/master/car_instance/car_models.py
For the red bbox in the picture below, its id in train.csv is 07, which corresponding to feiyate in the repo. But the car in the red bbox is obviously not a feiyate. Did I miss something here?
Also I had noticed that someone said we can get the car size(length, width and height) by the car model. Did he/she mean the average size? Or different size with different car?
\n\nOne last question.. Did the size of car model is calculated by following formula?\n(vector is in car_model.json)\nX_size = abs( max(vector[:,0]) - min(vector[:,0]) )\nY_size = abs( max(vector[:,1]) - min(vector[:,1]) )\nZ_size = abs( max(vector[:,2]) - min(vector[:,2]) )
I guess most of the people is regressing the rotation and translation information directly, or at least with some simple decoded. But there are a lots of different methods to predict the 3D information, like
\n\n2.Turn the rotation information into multi-bins for training and train with size of car for generating 3D bounding box in inference.
\n\nI think size information is needed or helpful in these methods. So I was wondering how to get the precise size information. Thanks in advances
", "messages": [ { "id": 708609, "postDate": "2020-01-02T13:35:54.143Z", "content": "I was trying to find the relationship between id in train.csv and the car model label in Apollo dataset api repo as following link:\nhttps://github.com/ApolloScapeAuto/dataset-api/blob/master/car_instance/car_models.py
For the red bbox in the picture below, its id in train.csv is 07, which corresponding to feiyate in the repo. But the car in the red bbox is obviously not a feiyate. Did I miss something here?
Also I had noticed that someone said we can get the car size(length, width and height) by the car model. Did he/she mean the average size? Or different size with different car?
\n\nOne last question.. Did the size of car model is calculated by following formula?\n(vector is in car_model.json)\nX_size = abs( max(vector[:,0]) - min(vector[:,0]) )\nY_size = abs( max(vector[:,1]) - min(vector[:,1]) )\nZ_size = abs( max(vector[:,2]) - min(vector[:,2]) )
I guess most of the people is regressing the rotation and translation information directly, or at least with some simple decoded. But there are a lots of different methods to predict the 3D information, like
\n\n2.Turn the rotation information into multi-bins for training and train with size of car for generating 3D bounding box in inference.
\n\nI think size information is needed or helpful in these methods. So I was wondering how to get the precise size information. Thanks in advances
", "rawMarkdown": "I was trying to find the relationship between `id` in train.csv and the car model label in Apollo dataset api repo as following link:\nhttps://github.com/ApolloScapeAuto/dataset-api/blob/master/car_instance/car_models.py\n\nFor the red bbox in the picture below, its `id` in train.csv is 07, which corresponding to `feiyate` in the repo. But the car in the red bbox is obviously not a `feiyate`. Did I miss something here? \n\n\n\nAlso I had noticed that someone said we can get the car size(length, width and height) by the car model. Did he/she mean the average size? Or different size with different car?\n\nOne last question.. Did the size of car model is calculated by following formula?\n(vector is in car_model.json)\n`X_size = abs( max(vector[:,0]) - min(vector[:,0]) )`\n`Y_size = abs( max(vector[:,1]) - min(vector[:,1]) )`\n`Z_size = abs( max(vector[:,2]) - min(vector[:,2]) )`\n\nI guess most of the people is regressing the rotation and translation information directly, or at least with some simple decoded. But there are a lots of different methods to predict the 3D information, like\n\n1. Predict the image points of 3D bounding box then use pnp to get the translation/rotation information.\n\n2.Turn the rotation information into multi-bins for training and train with size of car for generating 3D bounding box in inference.\n\nI think size information is needed or helpful in these methods. So I was wondering how to get the precise size information. Thanks in advances", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "708609": "I was trying to find the relationship between `id` in train.csv and the car model label in Apollo dataset api repo as following link:\nhttps://github.com/ApolloScapeAuto/dataset-api/blob/master/car_instance/car_models.py\n\nFor the red bbox in the picture below, its `id` in train.csv is 07, which corresponding to `feiyate` in the repo. But the car in the red bbox is obviously not a `feiyate`. Did I miss something here? \n\n\n\nAlso I had noticed that someone said we can get the car size(length, width and height) by the car model. Did he/she mean the average size? Or different size with different car?\n\nOne last question.. Did the size of car model is calculated by following formula?\n(vector is in car_model.json)\n`X_size = abs( max(vector[:,0]) - min(vector[:,0]) )`\n`Y_size = abs( max(vector[:,1]) - min(vector[:,1]) )`\n`Z_size = abs( max(vector[:,2]) - min(vector[:,2]) )`\n\nI guess most of the people is regressing the rotation and translation information directly, or at least with some simple decoded. But there are a lots of different methods to predict the 3D information, like\n\n1. Predict the image points of 3D bounding box then use pnp to get the translation/rotation information.\n\n2.Turn the rotation information into multi-bins for training and train with size of car for generating 3D bounding box in inference.\n\nI think size information is needed or helpful in these methods. So I was wondering how to get the precise size information. Thanks in advances" } }