{ "id": 188792, "title": "What about agent misclassifications?", "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/188792", "author_name": "", "post_date": "2020-10-05T10:29:17.578158Z", "votes": 10, "comment_count": 18, "views": 0, "content": "

Eyeballing the data, there seems to be many cases where an agent 'bounces' between locations. While some of these could be legitimate (cars parking/reversing, pedestrians alternating between locations), many seem to be more likely agent classification error. i.e. we think two distinct agents are the same and therefore measure movements between their locations as a single trajectory.

\n

Some examples below:
\n\"\"
\n\"\"
\n\"\"

\n

EDIT: thanks to @zaharch comment below I have now excluded non-available points from the trajectories. This cleans things up - we don't now see the same behaviour. There are some noteworthy patterns that seem to be due to noise that is unavoidable (I assume due to measurement/rasterization error of near-stationary agents). Some samples of these from the validation dataset (with timestamp - track_id) are below:

\n

\"\"
\n\"\"

", "messages": [ { "id": "1037775", "postDate": "10/05/2020 10:29:17", "content": "

Eyeballing the data, there seems to be many cases where an agent 'bounces' between locations. While some of these could be legitimate (cars parking/reversing, pedestrians alternating between locations), many seem to be more likely agent classification error. i.e. we think two distinct agents are the same and therefore measure movements between their locations as a single trajectory.

\n

Some examples below:
\n\"\"
\n\"\"
\n\"\"

\n

EDIT: thanks to @zaharch comment below I have now excluded non-available points from the trajectories. This cleans things up - we don't now see the same behaviour. There are some noteworthy patterns that seem to be due to noise that is unavoidable (I assume due to measurement/rasterization error of near-stationary agents). Some samples of these from the validation dataset (with timestamp - track_id) are below:

\n

\"\"
\n\"\"

", "rawMarkdown": "Eyeballing the data, there seems to be many cases where an agent 'bounces' between locations. While some of these could be legitimate (cars parking/reversing, pedestrians alternating between locations), many seem to be more likely agent classification error. i.e. we think two distinct agents are the same and therefore measure movements between their locations as a single trajectory.\n\nSome examples below:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F86cd9fb8af5cc0ce2bb8d08454f4b925%2Fsample4.jpeg?generation=1601893463421760&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2Ff786c97e3798ed8f0d27b13fb24516e8%2Fsample3.png?generation=1601893528468935&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F481f015ac5b2efce9e2146a2173cdf42%2Fsample2.png?generation=1601893541214503&alt=media)\n\n\nEDIT: thanks to @zaharch comment below I have now excluded non-available points from the trajectories. This cleans things up - we don't now see the same behaviour. There are some noteworthy patterns that seem to be due to noise that is unavoidable (I assume due to measurement/rasterization error of near-stationary agents). Some samples of these from the validation dataset (with timestamp - track_id) are below:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F8c386c6beec7b2fb07db9d3258ea47de%2Fsample_5.jpeg?generation=1601901770341363&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F3c21fae0c3bf4b5d8d723927e5a4f014%2Fsample_4.jpeg?generation=1601901788194282&alt=media)", "votes": null }, { "id": "1037842", "postDate": "10/05/2020 11:31:23", "content": "

Hi @fergusoci , it is interesting, can you give the ids? It seems that in some cases the graph jumps back to origin, and this can happen when availability is zero, coordinates become zero too. You need to filter those out before plotting.

", "rawMarkdown": "Hi @fergusoci , it is interesting, can you give the ids? It seems that in some cases the graph jumps back to origin, and this can happen when availability is zero, coordinates become zero too. You need to filter those out before plotting.", "votes": null }, { "id": "1037900", "postDate": "10/05/2020 12:23:38", "content": "

Very good point! I hadn't registered that when I ran through it this morning. I'm going to take another look now and pull back some ids.

", "rawMarkdown": "Very good point! I hadn't registered that when I ran through it this morning. I'm going to take another look now and pull back some ids.", "votes": null }, { "id": "1037947", "postDate": "10/05/2020 12:45:50", "content": "

That deals with those 'bouncing' patterns. I've edited the post accordingly. The ones that seem to be left now look to be mostly to do with measurement error, which I don't think we can do much to avoid. It shouldn't affect the nll metric too much in any case.

", "rawMarkdown": "That deals with those 'bouncing' patterns. I've edited the post accordingly. The ones that seem to be left now look to be mostly to do with measurement error, which I don't think we can do much to avoid. It shouldn't affect the nll metric too much in any case.", "votes": null }, { "id": "1037977", "postDate": "10/05/2020 13:05:20", "content": "

Yeah when plotting stationary objects it's clear that there is some amount of error in the trajectories.
\nIt's not so clear for fast moving agents, as a little bit of wiggle is probably much less than their motion, but if you plot the trajectories of \"stationary\" agents it's pretty clear how large this uncertainty is.

\n

I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.

\n

I though about using something to smooth out these trajectories (Kalman filter), but then again, our targets will have the same uncertainties, so not sure it will do much on average.

", "rawMarkdown": "Yeah when plotting stationary objects it's clear that there is some amount of error in the trajectories. \nIt's not so clear for fast moving agents, as a little bit of wiggle is probably much less than their motion, but if you plot the trajectories of \"stationary\" agents it's pretty clear how large this uncertainty is. \n\nI'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is. \n\nI though about using something to smooth out these trajectories (Kalman filter), but then again, our targets will have the same uncertainties, so not sure it will do much on average.", "votes": null }, { "id": "1039067", "postDate": "10/06/2020 09:44:38", "content": "
\n

I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.

\n
\n

I would expect it to exist. As a side note, one of the filters we run for agents is distance from ego, which addresses that (removing agents which are too far away and therefore too unreliable)

", "rawMarkdown": "> I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.\n\nI would expect it to exist. As a side note, one of the filters we run for agents is `distance from ego`, which addresses that (removing agents which are too far away and therefore too unreliable)", "votes": null }, { "id": "1043905", "postDate": "10/09/2020 10:21:11", "content": "

@lucabergamini Thanks for the info, it makes sense.
\nJust as an aside, what are the UNKNOWN agents generally?

\n

Are they just objects which might exist but probably aren't relevant (like trees, fire hydrants, etc.) or maybe objects which weren't in frame long enough to be properly classified?

\n

Also, I know only 4 of the possible 17 types of agents are found in the dataset. How were these filtered out? Did you specifically only choose scenes which only contained these 4 types of agents, or were they just dumped into UNKNOWN?

", "rawMarkdown": "lucabergamini Thanks for the info, it makes sense. \nJust as an aside, what are the UNKNOWN agents generally? \n\nAre they just objects which might exist but probably aren't relevant (like trees, fire hydrants, etc.) or maybe objects which weren't in frame long enough to be properly classified? \n\nAlso, I know only 4 of the possible 17 types of agents are found in the dataset. How were these filtered out? Did you specifically only choose scenes which only contained these 4 types of agents, or were they just dumped into UNKNOWN?", "votes": null }, { "id": "1046745", "postDate": "10/12/2020 00:58:17", "content": "

hello, great insight. I remember you raised a discussion about how to correct the agent trajectory. But I could not find it now. Can you share the link? Thanks.

", "rawMarkdown": "hello, great insight. I remember you raised a discussion about how to correct the agent trajectory. But I could not find it now. Can you share the link? Thanks.", "votes": null }, { "id": "1047064", "postDate": "10/12/2020 08:24:25", "content": "

Hi @shuozhang, that wasn't me, that was @zaharch here. If you're using the latest version of l5kit they have corrected it now. You should follow their sample .

", "rawMarkdown": "Hi @shuozhang, that wasn't me, that was @zaharch [here](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186492). If you're using the latest version of l5kit they have corrected it now. You should follow their [sample](https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb) .", "votes": null }, { "id": "1050893", "postDate": "10/15/2020 21:11:59", "content": "

There are definitely plenty of examples of this even in the chopped dataset where I presume the paths are fully available. I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason and others where there is a bit of squirrely-ness and clear discontinuity. It is definitely present but definitely the minority.

\n

This plot shows a scatter plot of my predictions in various colors and then the labeled path in blue. Path opacity represents confidence. Size of point represents time step (i.e. point 0 is very small and point 50 is the largest)

\n

\"\"

", "rawMarkdown": "There are definitely plenty of examples of this even in the chopped dataset where I presume the paths are fully available. I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason and others where there is a bit of squirrely-ness and clear discontinuity. It is definitely present but definitely the minority. \n\nThis plot shows a scatter plot of my predictions in various colors and then the labeled path in blue. Path opacity represents confidence. Size of point represents time step (i.e. point 0 is very small and point 50 is the largest)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1035002%2F8659701495d092cdb6d793e1ac8ed3ca%2Finconsistent.png?generation=1602796165841718&alt=media)", "votes": null }, { "id": "1050922", "postDate": "10/15/2020 22:35:32", "content": "

From the velocity this seems to be a pedestrian, so I wouldn't be too surprised about a sudden movement change. Did you check the corresponding satellite / semantic map if the GT path makes any sense?

", "rawMarkdown": "From the velocity this seems to be a pedestrian, so I wouldn't be too surprised about a sudden movement change. Did you check the corresponding satellite / semantic map if the GT path makes any sense?", "votes": null }, { "id": "1051183", "postDate": "10/16/2020 08:42:57", "content": "
\n

I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason

\n
\n

Are you sure the path is fully available? I mean, look at the target availabilities, as the \"filler\" value for when no target is found is 0, so maybe that could be confused for returning to the origin.

\n

But I agree with @ilu000 , pedestrian trajectories can be pretty wild. My guess again this would mostly be due to the fact that they move slowly, and therefore the relative uncertainty in their position is larger than that for vehicles.

", "rawMarkdown": "> I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason\n\nAre you sure the path is fully available? I mean, look at the target availabilities, as the \"filler\" value for when no target is found is 0, so maybe that could be confused for returning to the origin.\n\nBut I agree with @ilu000 , pedestrian trajectories can be pretty wild. My guess again this would mostly be due to the fact that they move slowly, and therefore the relative uncertainty in their position is larger than that for vehicles.", "votes": null }, { "id": "1051397", "postDate": "10/16/2020 13:33:59", "content": "

The pedestrian thing makes sense. Maybe my assumption is wrong but shouldn't the validation set created by the chopped dataset have full 50 timestep availability?

", "rawMarkdown": "The pedestrian thing makes sense. Maybe my assumption is wrong but shouldn't the validation set created by the chopped dataset have full 50 timestep availability?", "votes": null }, { "id": "1051407", "postDate": "10/16/2020 13:40:15", "content": "

I think it's 10 (min_frame_future = 10 in create_chopped_dataset())

", "rawMarkdown": "I think it's 10 (min_frame_future = 10 in create_chopped_dataset())", "votes": null }, { "id": "1051460", "postDate": "10/16/2020 14:28:55", "content": "

Yup, as @fergusoci said, it will return 50, but min is 10, so can be padded for up to 40 frames.
\nIf I don't mask my trajectories, you can see it \"jumps\" back to 0.
\n\"\"

\n

If I properly mask it:
\n\"\"

", "rawMarkdown": "Yup, as @fergusoci said, it will return 50, but min is 10, so can be padded for up to 40 frames. \nIf I don't mask my trajectories, you can see it \"jumps\" back to 0. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F314368%2F1981b665da150376b8700c17440e8447%2Fnomask.png?generation=1602858378321867&alt=media)\n\nIf I properly mask it: \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F314368%2F2ae6a5e2632e97085e567e4951c51536%2Fmask.png?generation=1602858525317078&alt=media)", "votes": null }, { "id": "1051498", "postDate": "10/16/2020 15:14:49", "content": "

Ahh I knew that. I saw it was set to 10 a while ago. It shows up very rarely in my data though. Kind of surprised I don't see it more often.

", "rawMarkdown": "Ahh I knew that. I saw it was set to 10 a while ago. It shows up very rarely in my data though. Kind of surprised I don't see it more often.", "votes": null }, { "id": "1055851", "postDate": "10/21/2020 06:53:31", "content": "

@fergusoci Hey :) I'm trying to reach you guys via an email from Kaggle. Could you check your inbox or maybe spam folder?

", "rawMarkdown": "fergusoci Hey :) I'm trying to reach you guys via an email from Kaggle. Could you check your inbox or maybe spam folder?", "votes": null }, { "id": "1056144", "postDate": "10/21/2020 13:06:23", "content": "

Hi Ivan, I haven't received anything, nothing in spam, etc. Maybe some issue on Kaggle side.

", "rawMarkdown": "Hi Ivan, I haven't received anything, nothing in spam, etc. Maybe some issue on Kaggle side.", "votes": null }, { "id": "1056359", "postDate": "10/21/2020 16:17:14", "content": "

That's odd. Could you send me anything via email, so that I can see yours? ivan.panshin@protonmail.com

", "rawMarkdown": "That's odd. Could you send me anything via email, so that I can see yours? ivan.panshin@protonmail.com", "votes": null } ], "comments": [ { "id": 1037842, "author_name": "zaharch", "author_url": "", "post_date": "10/05/2020 11:31:23", "content": "

Hi @fergusoci , it is interesting, can you give the ids? It seems that in some cases the graph jumps back to origin, and this can happen when availability is zero, coordinates become zero too. You need to filter those out before plotting.

", "votes": null, "replies": [ { "id": 1037900, "author_name": "fergusoci", "author_url": "", "post_date": "10/05/2020 12:23:38", "content": "

Very good point! I hadn't registered that when I ran through it this morning. I'm going to take another look now and pull back some ids.

", "votes": null, "replies": [] }, { "id": 1037947, "author_name": "fergusoci", "author_url": "", "post_date": "10/05/2020 12:45:50", "content": "

That deals with those 'bouncing' patterns. I've edited the post accordingly. The ones that seem to be left now look to be mostly to do with measurement error, which I don't think we can do much to avoid. It shouldn't affect the nll metric too much in any case.

", "votes": null, "replies": [] } ] }, { "id": 1037977, "author_name": "fnands", "author_url": "", "post_date": "10/05/2020 13:05:20", "content": "

Yeah when plotting stationary objects it's clear that there is some amount of error in the trajectories.
\nIt's not so clear for fast moving agents, as a little bit of wiggle is probably much less than their motion, but if you plot the trajectories of \"stationary\" agents it's pretty clear how large this uncertainty is.

\n

I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.

\n

I though about using something to smooth out these trajectories (Kalman filter), but then again, our targets will have the same uncertainties, so not sure it will do much on average.

", "votes": null, "replies": [ { "id": 1039067, "author_name": "lucabergamini", "author_url": "", "post_date": "10/06/2020 09:44:38", "content": "
\n

I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.

\n
\n

I would expect it to exist. As a side note, one of the filters we run for agents is distance from ego, which addresses that (removing agents which are too far away and therefore too unreliable)

", "votes": null, "replies": [] }, { "id": 1043905, "author_name": "fnands", "author_url": "", "post_date": "10/09/2020 10:21:11", "content": "

@lucabergamini Thanks for the info, it makes sense.
\nJust as an aside, what are the UNKNOWN agents generally?

\n

Are they just objects which might exist but probably aren't relevant (like trees, fire hydrants, etc.) or maybe objects which weren't in frame long enough to be properly classified?

\n

Also, I know only 4 of the possible 17 types of agents are found in the dataset. How were these filtered out? Did you specifically only choose scenes which only contained these 4 types of agents, or were they just dumped into UNKNOWN?

", "votes": null, "replies": [] } ] }, { "id": 1046745, "author_name": "shuozhang", "author_url": "", "post_date": "10/12/2020 00:58:17", "content": "

hello, great insight. I remember you raised a discussion about how to correct the agent trajectory. But I could not find it now. Can you share the link? Thanks.

", "votes": null, "replies": [ { "id": 1047064, "author_name": "fergusoci", "author_url": "", "post_date": "10/12/2020 08:24:25", "content": "

Hi @shuozhang, that wasn't me, that was @zaharch here. If you're using the latest version of l5kit they have corrected it now. You should follow their sample .

", "votes": null, "replies": [] }, { "id": 1055851, "author_name": "ivanpan", "author_url": "", "post_date": "10/21/2020 06:53:31", "content": "

@fergusoci Hey :) I'm trying to reach you guys via an email from Kaggle. Could you check your inbox or maybe spam folder?

", "votes": null, "replies": [] }, { "id": 1056144, "author_name": "fergusoci", "author_url": "", "post_date": "10/21/2020 13:06:23", "content": "

Hi Ivan, I haven't received anything, nothing in spam, etc. Maybe some issue on Kaggle side.

", "votes": null, "replies": [] }, { "id": 1056359, "author_name": "ivanpan", "author_url": "", "post_date": "10/21/2020 16:17:14", "content": "

That's odd. Could you send me anything via email, so that I can see yours? ivan.panshin@protonmail.com

", "votes": null, "replies": [] } ] }, { "id": 1050893, "author_name": "ryches", "author_url": "", "post_date": "10/15/2020 21:11:59", "content": "

There are definitely plenty of examples of this even in the chopped dataset where I presume the paths are fully available. I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason and others where there is a bit of squirrely-ness and clear discontinuity. It is definitely present but definitely the minority.

\n

This plot shows a scatter plot of my predictions in various colors and then the labeled path in blue. Path opacity represents confidence. Size of point represents time step (i.e. point 0 is very small and point 50 is the largest)

\n

\"\"

", "votes": null, "replies": [ { "id": 1050922, "author_name": "ilu000", "author_url": "", "post_date": "10/15/2020 22:35:32", "content": "

From the velocity this seems to be a pedestrian, so I wouldn't be too surprised about a sudden movement change. Did you check the corresponding satellite / semantic map if the GT path makes any sense?

", "votes": null, "replies": [] }, { "id": 1051183, "author_name": "fnands", "author_url": "", "post_date": "10/16/2020 08:42:57", "content": "
\n

I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason

\n
\n

Are you sure the path is fully available? I mean, look at the target availabilities, as the \"filler\" value for when no target is found is 0, so maybe that could be confused for returning to the origin.

\n

But I agree with @ilu000 , pedestrian trajectories can be pretty wild. My guess again this would mostly be due to the fact that they move slowly, and therefore the relative uncertainty in their position is larger than that for vehicles.

", "votes": null, "replies": [] }, { "id": 1051397, "author_name": "ryches", "author_url": "", "post_date": "10/16/2020 13:33:59", "content": "

The pedestrian thing makes sense. Maybe my assumption is wrong but shouldn't the validation set created by the chopped dataset have full 50 timestep availability?

", "votes": null, "replies": [] }, { "id": 1051407, "author_name": "fergusoci", "author_url": "", "post_date": "10/16/2020 13:40:15", "content": "

I think it's 10 (min_frame_future = 10 in create_chopped_dataset())

", "votes": null, "replies": [] }, { "id": 1051460, "author_name": "fnands", "author_url": "", "post_date": "10/16/2020 14:28:55", "content": "

Yup, as @fergusoci said, it will return 50, but min is 10, so can be padded for up to 40 frames.
\nIf I don't mask my trajectories, you can see it \"jumps\" back to 0.
\n\"\"

\n

If I properly mask it:
\n\"\"

", "votes": null, "replies": [] }, { "id": 1051498, "author_name": "ryches", "author_url": "", "post_date": "10/16/2020 15:14:49", "content": "

Ahh I knew that. I saw it was set to 10 a while ago. It shows up very rarely in my data though. Kind of surprised I don't see it more often.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1037775": "Eyeballing the data, there seems to be many cases where an agent 'bounces' between locations. While some of these could be legitimate (cars parking/reversing, pedestrians alternating between locations), many seem to be more likely agent classification error. i.e. we think two distinct agents are the same and therefore measure movements between their locations as a single trajectory.\n\nSome examples below:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F86cd9fb8af5cc0ce2bb8d08454f4b925%2Fsample4.jpeg?generation=1601893463421760&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2Ff786c97e3798ed8f0d27b13fb24516e8%2Fsample3.png?generation=1601893528468935&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F481f015ac5b2efce9e2146a2173cdf42%2Fsample2.png?generation=1601893541214503&alt=media)\n\n\nEDIT: thanks to @zaharch comment below I have now excluded non-available points from the trajectories. This cleans things up - we don't now see the same behaviour. There are some noteworthy patterns that seem to be due to noise that is unavoidable (I assume due to measurement/rasterization error of near-stationary agents). Some samples of these from the validation dataset (with timestamp - track_id) are below:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F8c386c6beec7b2fb07db9d3258ea47de%2Fsample_5.jpeg?generation=1601901770341363&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621560%2F3c21fae0c3bf4b5d8d723927e5a4f014%2Fsample_4.jpeg?generation=1601901788194282&alt=media)", "1037842": "Hi @fergusoci , it is interesting, can you give the ids? It seems that in some cases the graph jumps back to origin, and this can happen when availability is zero, coordinates become zero too. You need to filter those out before plotting.", "1037900": "Very good point! I hadn't registered that when I ran through it this morning. I'm going to take another look now and pull back some ids.", "1037947": "That deals with those 'bouncing' patterns. I've edited the post accordingly. The ones that seem to be left now look to be mostly to do with measurement error, which I don't think we can do much to avoid. It shouldn't affect the nll metric too much in any case.", "1037977": "Yeah when plotting stationary objects it's clear that there is some amount of error in the trajectories. \nIt's not so clear for fast moving agents, as a little bit of wiggle is probably much less than their motion, but if you plot the trajectories of \"stationary\" agents it's pretty clear how large this uncertainty is. \n\nI'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is. \n\nI though about using something to smooth out these trajectories (Kalman filter), but then again, our targets will have the same uncertainties, so not sure it will do much on average.", "1039067": "> I'd be curious to see if there is a correlation between how far an agent is from the ego and how large it's uncertainty is.\n\nI would expect it to exist. As a side note, one of the filters we run for agents is `distance from ego`, which addresses that (removing agents which are too far away and therefore too unreliable)", "1043905": "lucabergamini Thanks for the info, it makes sense. \nJust as an aside, what are the UNKNOWN agents generally? \n\nAre they just objects which might exist but probably aren't relevant (like trees, fire hydrants, etc.) or maybe objects which weren't in frame long enough to be properly classified? \n\nAlso, I know only 4 of the possible 17 types of agents are found in the dataset. How were these filtered out? Did you specifically only choose scenes which only contained these 4 types of agents, or were they just dumped into UNKNOWN?", "1046745": "hello, great insight. I remember you raised a discussion about how to correct the agent trajectory. But I could not find it now. Can you share the link? Thanks.", "1047064": "Hi @shuozhang, that wasn't me, that was @zaharch [here](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186492). If you're using the latest version of l5kit they have corrected it now. You should follow their [sample](https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb) .", "1050893": "There are definitely plenty of examples of this even in the chopped dataset where I presume the paths are fully available. I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason and others where there is a bit of squirrely-ness and clear discontinuity. It is definitely present but definitely the minority. \n\nThis plot shows a scatter plot of my predictions in various colors and then the labeled path in blue. Path opacity represents confidence. Size of point represents time step (i.e. point 0 is very small and point 50 is the largest)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1035002%2F8659701495d092cdb6d793e1ac8ed3ca%2Finconsistent.png?generation=1602796165841718&alt=media)", "1050922": "From the velocity this seems to be a pedestrian, so I wouldn't be too surprised about a sudden movement change. Did you check the corresponding satellite / semantic map if the GT path makes any sense?", "1051183": "> I found some where the agent goes in one direction in a straight line but then the final point is back at the origin for some reason\n\nAre you sure the path is fully available? I mean, look at the target availabilities, as the \"filler\" value for when no target is found is 0, so maybe that could be confused for returning to the origin.\n\nBut I agree with @ilu000 , pedestrian trajectories can be pretty wild. My guess again this would mostly be due to the fact that they move slowly, and therefore the relative uncertainty in their position is larger than that for vehicles.", "1051397": "The pedestrian thing makes sense. Maybe my assumption is wrong but shouldn't the validation set created by the chopped dataset have full 50 timestep availability?", "1051407": "I think it's 10 (min_frame_future = 10 in create_chopped_dataset())", "1051460": "Yup, as @fergusoci said, it will return 50, but min is 10, so can be padded for up to 40 frames. \nIf I don't mask my trajectories, you can see it \"jumps\" back to 0. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F314368%2F1981b665da150376b8700c17440e8447%2Fnomask.png?generation=1602858378321867&alt=media)\n\nIf I properly mask it: \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F314368%2F2ae6a5e2632e97085e567e4951c51536%2Fmask.png?generation=1602858525317078&alt=media)", "1051498": "Ahh I knew that. I saw it was set to 10 a while ago. It shows up very rarely in my data though. Kind of surprised I don't see it more often.", "1055851": "fergusoci Hey :) I'm trying to reach you guys via an email from Kaggle. Could you check your inbox or maybe spam folder?", "1056144": "Hi Ivan, I haven't received anything, nothing in spam, etc. Maybe some issue on Kaggle side.", "1056359": "That's odd. Could you send me anything via email, so that I can see yours? ivan.panshin@protonmail.com" }, "source": "meta" }