Since the competition is over I want to ask what the top teams in the leader board have done? As far as I know Rasterization is considered as bottle neck and most of the people have made many discussions on improving the training time while keeping the base model same (ResNet). I want to know if anyone has done something different algorithm wise and had success?
\nWe have LaneGCN or CS-LSTM which were successful for some other datasets. Does it make sense to use any of those algorithms or do we have to come up with something completely new this time?
", "messages": [ { "id": "1096777", "postDate": "11/30/2020 19:08:32", "content": "Since the competition is over I want to ask what the top teams in the leader board have done? As far as I know Rasterization is considered as bottle neck and most of the people have made many discussions on improving the training time while keeping the base model same (ResNet). I want to know if anyone has done something different algorithm wise and had success?
\nWe have LaneGCN or CS-LSTM which were successful for some other datasets. Does it make sense to use any of those algorithms or do we have to come up with something completely new this time?
", "rawMarkdown": "Since the competition is over I want to ask what the top teams in the leader board have done? As far as I know Rasterization is considered as bottle neck and most of the people have made many discussions on improving the training time while keeping the base model same (ResNet). I want to know if anyone has done something different algorithm wise and had success? \n\nWe have LaneGCN or CS-LSTM which were successful for some other datasets. Does it make sense to use any of those algorithms or do we have to come up with something completely new this time?", "votes": null }, { "id": "1098653", "postDate": "12/01/2020 18:58:57", "content": "There is one using vectornet (10th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199711.
\nAnd, there is also mixnet (9th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199636
\nThe 3rd place use xception41 https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199499#1091376
Thanks a lot for the comment. So basically 9th and 3rd that you mentioned used a feature extractor and did some data cleaning and changed the training procedure. The 10th place submission is the only one that has used something that is particular to motion prediction.
", "rawMarkdown": "Thanks a lot for the comment. So basically 9th and 3rd that you mentioned used a feature extractor and did some data cleaning and changed the training procedure. The 10th place submission is the only one that has used something that is particular to motion prediction.", "votes": null } ], "comments": [ { "id": 1098653, "author_name": "louis925", "author_url": "", "post_date": "12/01/2020 18:58:57", "content": "There is one using vectornet (10th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199711.
\nAnd, there is also mixnet (9th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199636
\nThe 3rd place use xception41 https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199499#1091376
Thanks a lot for the comment. So basically 9th and 3rd that you mentioned used a feature extractor and did some data cleaning and changed the training procedure. The 10th place submission is the only one that has used something that is particular to motion prediction.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1096777": "Since the competition is over I want to ask what the top teams in the leader board have done? As far as I know Rasterization is considered as bottle neck and most of the people have made many discussions on improving the training time while keeping the base model same (ResNet). I want to know if anyone has done something different algorithm wise and had success? \n\nWe have LaneGCN or CS-LSTM which were successful for some other datasets. Does it make sense to use any of those algorithms or do we have to come up with something completely new this time?", "1098653": "There is one using vectornet (10th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199711. \nAnd, there is also mixnet (9th place solution) https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199636\nThe 3rd place use xception41 https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/199499#1091376", "1099325": "Thanks a lot for the comment. So basically 9th and 3rd that you mentioned used a feature extractor and did some data cleaning and changed the training procedure. The 10th place submission is the only one that has used something that is particular to motion prediction." }, "source": "meta" }