\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1010129%2F851ac7c955256b166ec2357de7f56d4a%2F__results___20_0.png?generation=1604654601973829&alt=media\")
I made a notebook to visualize your submission.csv. Most top rank people probably had already done this but here is the notebook in case you haven't figured that out:
\nhttps://www.kaggle.com/louis925/lyft-submission-visualization
Example plots from the public 23.622 submission
\n
I made a notebook to visualize your submission.csv. Most top rank people probably had already done this but here is the notebook in case you haven't figured that out:
\nhttps://www.kaggle.com/louis925/lyft-submission-visualization
Example plots from the public 23.622 submission
\n
Nice work, thanks for sharing.
", "rawMarkdown": "Nice work, thanks for sharing.", "votes": null }, { "id": "1070921", "postDate": "11/06/2020 10:12:04", "content": "great job!
", "rawMarkdown": "great job!", "votes": null }, { "id": "1071083", "postDate": "11/06/2020 13:33:09", "content": "Great job! Thanks!!!!
", "rawMarkdown": "Great job! Thanks!!!!", "votes": null }, { "id": "1071238", "postDate": "11/06/2020 16:38:02", "content": "have you seen jitter in trajectories with high nnl score especially where previous or future road is cut off from semantic map? for example.
\nnnl score = 6640 , confidence of blue =99%
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4573165%2F6f7c705ec738257072be434852f17099%2Fdownload.png?generation=1604681247888641&alt=media\")
Not really but looks like your transformation after the prediction is not correct. I have done an experiment and found that constant prediction can give your score about ~6000. So anything worse than that can mean something wrong with the transformation. Also, looks like your future trajectory doesn't start from your agent current location.
", "rawMarkdown": "Not really but looks like your transformation after the prediction is not correct. I have done an experiment and found that constant prediction can give your score about ~6000. So anything worse than that can mean something wrong with the transformation. Also, looks like your future trajectory doesn't start from your agent current location.", "votes": null }, { "id": "1071480", "postDate": "11/06/2020 23:47:31", "content": "actually that score was just for that particular sample. I sorted the samples in validation data based on their nnl score in descending order and that example was amongst the top ones. for rest of the samples nnl is low. for example this one has score of 0.456
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4573165%2Fcf4f95d88b2ba2a877235542b4686e88%2Fdownload.png?generation=1604706336356530&alt=media\")
It looks like when road is not visible in the past cnn don't perform well.
", "rawMarkdown": "actually that score was just for that particular sample. I sorted the samples in validation data based on their nnl score in descending order and that example was amongst the top ones. for rest of the samples nnl is low. for example this one has score of 0.456\n\n\nIt looks like when road is not visible in the past cnn don't perform well.", "votes": null }, { "id": "1071580", "postDate": "11/07/2020 04:37:22", "content": "ok, yeah, I am not sure how you get the plots. But looks like all the trajectories are not start from the target agent (green) unless you label the ego rather than agent.
\nIf you are using the resnet model that available on the Kaggle, then right, once the road are cut off, the model has no information for the rest of the road. However, the pictures you show doesn't look to me has any cutoff. Their history and future trajectory are all within the plot.
Very useful topic and kernel, thank you
", "rawMarkdown": "Very useful topic and kernel, thank you", "votes": null } ], "comments": [ { "id": 1070899, "author_name": "mohaiminul101", "author_url": "", "post_date": "11/06/2020 09:39:49", "content": "Nice work, thanks for sharing.
", "votes": null, "replies": [] }, { "id": 1070921, "author_name": "maciejgronczynski", "author_url": "", "post_date": "11/06/2020 10:12:04", "content": "great job!
", "votes": null, "replies": [] }, { "id": 1071083, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "11/06/2020 13:33:09", "content": "Great job! Thanks!!!!
", "votes": null, "replies": [] }, { "id": 1071238, "author_name": "sujaydkhandekar", "author_url": "", "post_date": "11/06/2020 16:38:02", "content": "have you seen jitter in trajectories with high nnl score especially where previous or future road is cut off from semantic map? for example.
\nnnl score = 6640 , confidence of blue =99%
\n
Not really but looks like your transformation after the prediction is not correct. I have done an experiment and found that constant prediction can give your score about ~6000. So anything worse than that can mean something wrong with the transformation. Also, looks like your future trajectory doesn't start from your agent current location.
", "votes": null, "replies": [] }, { "id": 1071480, "author_name": "sujaydkhandekar", "author_url": "", "post_date": "11/06/2020 23:47:31", "content": "actually that score was just for that particular sample. I sorted the samples in validation data based on their nnl score in descending order and that example was amongst the top ones. for rest of the samples nnl is low. for example this one has score of 0.456
\n
It looks like when road is not visible in the past cnn don't perform well.
", "votes": null, "replies": [] }, { "id": 1071580, "author_name": "louis925", "author_url": "", "post_date": "11/07/2020 04:37:22", "content": "ok, yeah, I am not sure how you get the plots. But looks like all the trajectories are not start from the target agent (green) unless you label the ego rather than agent.
\nIf you are using the resnet model that available on the Kaggle, then right, once the road are cut off, the model has no information for the rest of the road. However, the pictures you show doesn't look to me has any cutoff. Their history and future trajectory are all within the plot.
Very useful topic and kernel, thank you
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1070893": "I made a notebook to visualize your `submission.csv`. Most top rank people probably had already done this but here is the notebook in case you haven't figured that out:\nhttps://www.kaggle.com/louis925/lyft-submission-visualization\n\nExample plots from the public 23.622 submission\n", "1070899": "Nice work, thanks for sharing.", "1070921": "great job!", "1071083": "Great job! Thanks!!!!", "1071238": "have you seen jitter in trajectories with high nnl score especially where previous or future road is cut off from semantic map? for example.\n\nnnl score = 6640 , confidence of blue =99%\n", "1071351": "Not really but looks like your transformation after the prediction is not correct. I have done an experiment and found that constant prediction can give your score about ~6000. So anything worse than that can mean something wrong with the transformation. Also, looks like your future trajectory doesn't start from your agent current location.", "1071480": "actually that score was just for that particular sample. I sorted the samples in validation data based on their nnl score in descending order and that example was amongst the top ones. for rest of the samples nnl is low. for example this one has score of 0.456\n\n\nIt looks like when road is not visible in the past cnn don't perform well.", "1071580": "ok, yeah, I am not sure how you get the plots. But looks like all the trajectories are not start from the target agent (green) unless you label the ego rather than agent.\nIf you are using the resnet model that available on the Kaggle, then right, once the road are cut off, the model has no information for the rest of the road. However, the pictures you show doesn't look to me has any cutoff. Their history and future trajectory are all within the plot.", "1073213": "Very useful topic and kernel, thank you" }, "source": "meta" }