Hey, other self-driving card perception competitions I've participated in impose some level of inference time limit (associated with a particular HW, e.g. 1080 Ti + x86 CPU w/ 8 cores @ 3.8 GHz etc) eg 100msec per lidar frame... so that winning solutions have a shot at being deployed in a car.
\n\nNot imposing this may still be beneficial to you even if inference is not real-time (help annotations) but from competition standpoint it will get us again into ensembling hell.
\n\nCurious about this lack of requirement.
", "messages": [ { "id": "625470", "postDate": "09/13/2019 05:42:43", "content": "Hey, other self-driving card perception competitions I've participated in impose some level of inference time limit (associated with a particular HW, e.g. 1080 Ti + x86 CPU w/ 8 cores @ 3.8 GHz etc) eg 100msec per lidar frame... so that winning solutions have a shot at being deployed in a car.
\n\nNot imposing this may still be beneficial to you even if inference is not real-time (help annotations) but from competition standpoint it will get us again into ensembling hell.
\n\nCurious about this lack of requirement.
", "rawMarkdown": "Hey, other self-driving card perception competitions I've participated in impose some level of inference time limit (associated with a particular HW, e.g. 1080 Ti + x86 CPU w/ 8 cores @ 3.8 GHz etc) eg 100msec per lidar frame... so that winning solutions have a shot at being deployed in a car.\n\nNot imposing this may still be beneficial to you even if inference is not real-time (help annotations) but from competition standpoint it will get us again into ensembling hell.\n\nCurious about this lack of requirement.", "votes": null }, { "id": "625658", "postDate": "09/13/2019 09:11:20", "content": "May be they didn't want to launch a new kernel (inference) competition in parallel, given the new massive limitation on Kernel GPU usage (The dataset is also huge).
\n\nBut you're right, for once a fast and efficient model is really necessary in practice, this may turn into ensembling hell.
", "rawMarkdown": "May be they didn't want to launch a new kernel (inference) competition in parallel, given the new massive limitation on Kernel GPU usage (The dataset is also huge). \n\nBut you're right, for once a fast and efficient model is really necessary in practice, this may turn into ensembling hell.", "votes": null }, { "id": "625932", "postDate": "09/13/2019 15:40:52", "content": "I’m also very curious, what kind of hardware comparison is on a SDC? (In terms of equivalence, can we equate it to a 1080Ti + 8 Core Server as @antorsae mentioned?)
\n\nIIRC in the AI Podcast, Elon Musk mentioned that they have 2 computer on board Tesla which aren’t even being used heavily when performing the autonomous functions.
", "rawMarkdown": "I’m also very curious, what kind of hardware comparison is on a SDC? (In terms of equivalence, can we equate it to a 1080Ti + 8 Core Server as @antorsae mentioned?)\n\nIIRC in the AI Podcast, Elon Musk mentioned that they have 2 computer on board Tesla which aren’t even being used heavily when performing the autonomous functions.", "votes": null }, { "id": "625963", "postDate": "09/13/2019 16:28:05", "content": "I think it is safe to assume that they use something like the NVIDIA DRIVE platform. Perhaps they use it, or something faster for real-time debugging.
\n\nNevertheless, the Tesla Autonomy day presentations give good insights into custom chipset development and therefore, we could guess what they are using as neural nets (even without LiDAR).
\n\nFinally, we shouldn't forget what comma.ai's openpilot is able to do on a quite restricted platform.
\n\nGenerally speaking, as soon as we go into ASICs, it is really difficult to compare performance to desktop or server systems because they cover a broader field of applications and therefore have much more overhead per Watt.
", "rawMarkdown": "I think it is safe to assume that they use something like the [NVIDIA DRIVE](https://www.nvidia.com/en-us/self-driving-cars/drive-platform/) platform. Perhaps they use it, or something faster for real-time debugging.\n\nNevertheless, the [Tesla Autonomy day presentations](https://youtu.be/Ucp0TTmvqOE?t=4516) give good insights into custom chipset development and therefore, we could guess what they are using as neural nets (even without LiDAR).\n\nFinally, we shouldn't forget what [comma.ai's openpilot](https://github.com/commaai/openpilot) is able to do on a quite restricted platform. \n\nGenerally speaking, as soon as we go into ASICs, it is really difficult to compare performance to desktop or server systems because they cover a broader field of applications and therefore have much more overhead per Watt.", "votes": null }, { "id": "626074", "postDate": "09/13/2019 18:56:31", "content": "I cannot recall any competition, where the winning solution went to production. We do not expect that the faith of winning solution in this challenge would be different. :)
\n\nAt the same time ideas and tricks that the community will develop will be extremely valuable.
\n\nI like the idea of competitions with limited resources. Especially if it is in the format of Docker containers, as it was at the SpaceNet challenge.
\n\nBut this type of data and this type of problem are new at Kaggle. It is not 2D image segmentation and detection which became common knowledge and nearly everyone knows how to approach it.
\n\nIn this challenge, we decided not to put the constraints on inference and make the participants work with the data in the way they like.
\n\nI am also very curious to see how people will create ensembles in 3D detection. :)
\n\nI am also not sure, that ensembling is the way to go here. The dataset is rich. There are so many interesting ways to combine the data from different sensors. I believe that in two months we will see something on the creative and not brute force side.
\n\nThe time will show :)
", "rawMarkdown": "I cannot recall any competition, where the winning solution went to production. We do not expect that the faith of winning solution in this challenge would be different. :)\n\nAt the same time ideas and tricks that the community will develop will be extremely valuable.\n\nI like the idea of competitions with limited resources. Especially if it is in the format of Docker containers, as it was at the SpaceNet challenge.\n\nBut this type of data and this type of problem are new at Kaggle. It is not 2D image segmentation and detection which became common knowledge and nearly everyone knows how to approach it.\n\nIn this challenge, we decided not to put the constraints on inference and make the participants work with the data in the way they like.\n\nI am also very curious to see how people will create ensembles in 3D detection. :)\n\nI am also not sure, that ensembling is the way to go here. The dataset is rich. There are so many interesting ways to combine the data from different sensors. I believe that in two months we will see something on the creative and not brute force side. \n\nThe time will show :)", "votes": null } ], "comments": [ { "id": 625658, "author_name": "serigne", "author_url": "", "post_date": "09/13/2019 09:11:20", "content": "May be they didn't want to launch a new kernel (inference) competition in parallel, given the new massive limitation on Kernel GPU usage (The dataset is also huge).
\n\nBut you're right, for once a fast and efficient model is really necessary in practice, this may turn into ensembling hell.
", "votes": null, "replies": [] }, { "id": 625932, "author_name": "init27", "author_url": "", "post_date": "09/13/2019 15:40:52", "content": "I’m also very curious, what kind of hardware comparison is on a SDC? (In terms of equivalence, can we equate it to a 1080Ti + 8 Core Server as @antorsae mentioned?)
\n\nIIRC in the AI Podcast, Elon Musk mentioned that they have 2 computer on board Tesla which aren’t even being used heavily when performing the autonomous functions.
", "votes": null, "replies": [ { "id": 625963, "author_name": "simonwenkel", "author_url": "", "post_date": "09/13/2019 16:28:05", "content": "I think it is safe to assume that they use something like the NVIDIA DRIVE platform. Perhaps they use it, or something faster for real-time debugging.
\n\nNevertheless, the Tesla Autonomy day presentations give good insights into custom chipset development and therefore, we could guess what they are using as neural nets (even without LiDAR).
\n\nFinally, we shouldn't forget what comma.ai's openpilot is able to do on a quite restricted platform.
\n\nGenerally speaking, as soon as we go into ASICs, it is really difficult to compare performance to desktop or server systems because they cover a broader field of applications and therefore have much more overhead per Watt.
", "votes": null, "replies": [] } ] }, { "id": 626074, "author_name": "iglovikov", "author_url": "", "post_date": "09/13/2019 18:56:31", "content": "I cannot recall any competition, where the winning solution went to production. We do not expect that the faith of winning solution in this challenge would be different. :)
\n\nAt the same time ideas and tricks that the community will develop will be extremely valuable.
\n\nI like the idea of competitions with limited resources. Especially if it is in the format of Docker containers, as it was at the SpaceNet challenge.
\n\nBut this type of data and this type of problem are new at Kaggle. It is not 2D image segmentation and detection which became common knowledge and nearly everyone knows how to approach it.
\n\nIn this challenge, we decided not to put the constraints on inference and make the participants work with the data in the way they like.
\n\nI am also very curious to see how people will create ensembles in 3D detection. :)
\n\nI am also not sure, that ensembling is the way to go here. The dataset is rich. There are so many interesting ways to combine the data from different sensors. I believe that in two months we will see something on the creative and not brute force side.
\n\nThe time will show :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "625470": "Hey, other self-driving card perception competitions I've participated in impose some level of inference time limit (associated with a particular HW, e.g. 1080 Ti + x86 CPU w/ 8 cores @ 3.8 GHz etc) eg 100msec per lidar frame... so that winning solutions have a shot at being deployed in a car.\n\nNot imposing this may still be beneficial to you even if inference is not real-time (help annotations) but from competition standpoint it will get us again into ensembling hell.\n\nCurious about this lack of requirement.", "625658": "May be they didn't want to launch a new kernel (inference) competition in parallel, given the new massive limitation on Kernel GPU usage (The dataset is also huge). \n\nBut you're right, for once a fast and efficient model is really necessary in practice, this may turn into ensembling hell.", "625932": "I’m also very curious, what kind of hardware comparison is on a SDC? (In terms of equivalence, can we equate it to a 1080Ti + 8 Core Server as @antorsae mentioned?)\n\nIIRC in the AI Podcast, Elon Musk mentioned that they have 2 computer on board Tesla which aren’t even being used heavily when performing the autonomous functions.", "625963": "I think it is safe to assume that they use something like the [NVIDIA DRIVE](https://www.nvidia.com/en-us/self-driving-cars/drive-platform/) platform. Perhaps they use it, or something faster for real-time debugging.\n\nNevertheless, the [Tesla Autonomy day presentations](https://youtu.be/Ucp0TTmvqOE?t=4516) give good insights into custom chipset development and therefore, we could guess what they are using as neural nets (even without LiDAR).\n\nFinally, we shouldn't forget what [comma.ai's openpilot](https://github.com/commaai/openpilot) is able to do on a quite restricted platform. \n\nGenerally speaking, as soon as we go into ASICs, it is really difficult to compare performance to desktop or server systems because they cover a broader field of applications and therefore have much more overhead per Watt.", "626074": "I cannot recall any competition, where the winning solution went to production. We do not expect that the faith of winning solution in this challenge would be different. :)\n\nAt the same time ideas and tricks that the community will develop will be extremely valuable.\n\nI like the idea of competitions with limited resources. Especially if it is in the format of Docker containers, as it was at the SpaceNet challenge.\n\nBut this type of data and this type of problem are new at Kaggle. It is not 2D image segmentation and detection which became common knowledge and nearly everyone knows how to approach it.\n\nIn this challenge, we decided not to put the constraints on inference and make the participants work with the data in the way they like.\n\nI am also very curious to see how people will create ensembles in 3D detection. :)\n\nI am also not sure, that ensembling is the way to go here. The dataset is rich. There are so many interesting ways to combine the data from different sensors. I believe that in two months we will see something on the creative and not brute force side. \n\nThe time will show :)" }, "source": "meta" }