Visualization of predicted trajectory

[abduallahmohamed]

Visualization of predicted trajectory

[abduallahmohamed]

In short for the visualization:
You figure out the homography matrix H for your dataset of interest.
Then you sample the video and capture a frame each 0.2 seconds, as mentioned in the dataset website like eth.
Then you do your predictions and transform them to the real world using H.
Last thing, to draw the distribution, it's a matter of choice, you can samples many trajectories from the predicted distribution and fit a KDE to it.

Hope this helps

[yeyingdege]

In short for the visualization:
You figure out the homography matrix H for your dataset of interest.
Then you sample the video and capture a frame each 0.2 seconds, as mentioned in the dataset website like eth.
Then you do your predictions and transform them to the real world using H.
Last thing, to draw the distribution, it's a matter of choice, you can samples many trajectories from the predicted distribution and fit a KDE to it.

Hope this helps

@abduallahmohamed Sorry to bother. What does KDE mean?

[xingchenzhang]

Hi @abduallahmohamed,

I also have some questions about the visualization.

1. You mentioned that ' sample the video and capture a frame each 0.2 seconds', I think this should be 0.4 seconds, right? because the coordiantes are sampled every 0.4 seconds

2. To obtain a predicted trajectory, we need to make a sample from the predicted distribution, right?

3. You mentioned 'Then you do your predictions and transform them to the real world using H'. It seems that the prediction we obtain is already read world coordinate? Maybe what we need to do is to trasnform them to pixel coordinate using H?

I am not sure if my understanding is correct...

Thank you very much for your help!
Bests,
Xingchen

[abduallahmohamed]

Hi Xingchen, Thanks for asking. 1. It should be 0.4 seconds, thanks for correction. 2. Yes, each predicted time step per pedestrian has a distribution that you need to sample from. 3. As you clarified, I meant the predictions are in real word (meters) but for visualization purposes you need to transform the results into image domain which uses H. Best, ------------------------------------------ Abduallah Mohamed abduallahmohamed.com

…

On Mon, Nov 2, 2020 at 7:16 AM Xingchen Zhang ***@***.***> wrote: Hi @abduallahmohamed <https://github.com/abduallahmohamed>, I also have some questions about the visualization. 1. You mentioned that ' sample the video and capture a frame each 0.2 seconds', I think this should be 0.4 seconds, right? because the coordiantes are sampled every 0.4 seconds 2. To obtain a predicted trajectory, we need to make a sample from the predicted distribution, right? 3. You mentioned 'Then you do your predictions and transform them to the real world using H'. It seems that the prediction we obtain is already read world coordinate? Maybe what we need to do is to trasnform them to pixel coordinate using H? I am not sure if my understanding is correct... Thank you very much for your help! Bests, Xingchen — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFIVLG6W7SVI2AJGHHV5EATSN2WMBANCNFSM4LCGS4JA> .

[xingchenzhang]

Hi Abduallah @abduallahmohamed,

Thank you very much for your reply. I am now trying to visualize the trajectory. I have sampled the videos, and also have the H matrix. However, I am wondering:

1. In your code, how you know which pedestrian and in which frame you are predicting? This is need to find the ground truth to do the plot.
2. Which format should I use to store the predicted trajectory, so that I can compare with other algorithm?

Thanks a lot!
Xingchen

[abduallahmohamed]

Hi Xingche, 1. If I recall correctly, I converted the video into frames using the known OpenCV code that converts video to frames, the videos as I recall were 30 fps. Then the sampling rate is 0.4 sec, so each 0.4 sec I chose the frame and align it [ In the data original files there's a column for frame number, this how you align it correctly] 2. I think each work uses it's own format, so you will have to write down a wrapper for each method predictions to convert it into your suitable format. Best, ------------------------------------------ Abduallah Mohamed abduallahmohamed.com

…

On Wed, Nov 4, 2020 at 3:57 PM Xingchen Zhang ***@***.***> wrote: Hi Abduallah @abduallahmohamed <https://github.com/abduallahmohamed>, Thank you very much for your reply. I am now trying to visualize the trajectory. I have sampled the videos, and also have the H matrix. However, I am wondering: 1. In your code, how you know which pedestrian and in which frame you are predicting? This is need to find the ground truth to do the plot. 2. Which format should I use to store the predicted trajectory, so that I can compare with other algorithm? Thanks a lot! Xingchen — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFIVLGZPBAGS4MK6OL6OV5LSOHE5RANCNFSM4LCGS4JA> .

[xingchenzhang]

Hi Abduallah @abduallahmohamed,

Thank you very much for your reply, it is very helpful. I have worked out these. Another thing I am doubting is how to deal with the negative values in the test file? For example hotel/test/biwi_hotel.txt?

Thanks a lot!

[abduallahmohamed]

Hi, What do you mean by negative values? To my understanding it’s when a pedestrian makes a turn or walks in the opposite direction. If so, there will be no issue as my module deals with delta of the pose and the activation function have a negative range.

On Sun, Nov 8, 2020 at 3:43 PM Xingchen Zhang ***@***.***> wrote: Hi Abduallah @abduallahmohamed <https://github.com/abduallahmohamed>, Thank you very much for your reply, it is very helpful. I have worked out these. Another thing I am doubting is how to deal with the negative values in the test file? For example hotel/test/biwi_hotel.txt? Thanks a lot! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFIVLG65LKS4YBFAWUNXWVDSO4GGTANCNFSM4LCGS4JA> .

-- ------------------------------------------ Abduallah Mohamed abduallahmohamed.com

[xingchenzhang]

Hi Abduallah @abduallahmohamed ,

Thank you very much for your reply. Regarding the negative value I mean in the training set and test set, some coordinates are negtive. But now I know it doesn't matter, and I have successfully plot trajectories. Thank you very much for your answers, which are very useful.

A final thing that I wish to ask, I don't know if you could give some suggestions according to your experience. In your code, after processing the test set, the frame ID and pedestrian ID are not preserved, only the (x,y) coordinates are used. Therefore the after sampling from the predicted distribution, we do not have the frame ID and pedestrian ID. Therefore, it is not very efficient to plot the trajectory. I am wondering do you have any suggestions to improve the efficiency of finding corresponding prediction results, frame ID and pedestrian ID.

Thanks a lot!

Bests,
Xingchen

[abduallahmohamed]

Hi Xingchen, Thanks for the discussion, it's really useful. I think you need the frame ID in order to visualize the results. What I did before is to add this piece of information in the dataloader itself: https://github.com/abduallahmohamed/Social-STGCNN/blob/master/utils.py#L125 I wish I had this code ready or had time to write it down. Anyways, if you have the visualization script done on you side and willing to share it, please do a merge request into the repo. Best, ------------------------------------------ Abduallah Mohamed abduallahmohamed.com

…

On Mon, Nov 9, 2020 at 8:23 AM Xingchen Zhang ***@***.***> wrote: Hi Abduallah @abduallahmohamed <https://github.com/abduallahmohamed> , Thank you very much for your reply. Regarding the negative value I mean in the training set and test set, some coordinates are negtive. But now I know it doesn't matter, and I have successfully plot trajectories. Thank you very much for your answers, which are very useful. A final thing that I wish to ask, I don't know if you could give some suggestions according to your experience. In your code, after processing the test set, the frame ID and pedestrian ID are not preserved, only the (x,y) coordinates are used. Therefore the after sampling from the predicted distribution, we do not have the frame ID and pedestrian ID. Therefore, it is not very efficient to plot the trajectory. I am wondering do you have any suggestions to improve the efficiency of finding corresponding prediction results, frame ID and pedestrian ID. Thanks a lot! Bests, Xingchen — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFIVLGYQPH2SF5BI3LCEAV3SO73OBANCNFSM4LCGS4JA> .

[henbucuoshanghai]

help where is the visualization script?in test.py?

[chuli97]

Hi Abduallah @abduallahmohamed ,

Thank you very much for your reply. Regarding the negative value I mean in the training set and test set, some coordinates are negtive. But now I know it doesn't matter, and I have successfully plot trajectories. Thank you very much for your answers, which are very useful.

A final thing that I wish to ask, I don't know if you could give some suggestions according to your experience. In your code, after processing the test set, the frame ID and pedestrian ID are not preserved, only the (x,y) coordinates are used. Therefore the after sampling from the predicted distribution, we do not have the frame ID and pedestrian ID. Therefore, it is not very efficient to plot the trajectory. I am wondering do you have any suggestions to improve the efficiency of finding corresponding prediction results, frame ID and pedestrian ID.

Thanks a lot!

Bests, Xingchen

嗨 Abduallah @abduallahmohamed，

非常感谢您的回复。关于我在训练集和测试集中的负值，一些坐标是负的。但现在我知道没关系，我已经成功地绘制了轨迹。非常感谢您的回答，非常有用。

最后想问一下，不知道各位能否根据自己的经验给点建议。在您的代码中，处理完测试集后，不保留框架 ID 和行人 ID，仅使用 (x,y) 坐标。因此，从预测分布中采样后，我们没有帧 ID 和行人 ID。因此，绘制轨迹不是很有效。我想知道您有什么建议可以提高查找相应预测结果、帧 ID 和行人 ID 的效率。

非常感谢！

最好的， 星辰

Hello，
I've been learning about trajectory prediction recently, and I want this project to predict the trajectory of my own video data. Can you share your visualization code with me?
If you can, please email me: [email protected] ,
be deeply grateful！

[abduallahmohamed]

Hi, To add the id, simply modify the dataset code while parsing to include pedestrians IDs starring from here: https://github.com/abduallahmohamed/Social-STGCNN/blob/master/utils.py#L122 For visualization: #58 ------------------------------------------ Abduallah Mohamed abduallahmohamed.com

…

On Mon, 25 Apr 2022 at 06:17, chuli97 ***@***.***> wrote: Hi Abduallah @abduallahmohamed <https://github.com/abduallahmohamed> , Thank you very much for your reply. Regarding the negative value I mean in the training set and test set, some coordinates are negtive. But now I know it doesn't matter, and I have successfully plot trajectories. Thank you very much for your answers, which are very useful. A final thing that I wish to ask, I don't know if you could give some suggestions according to your experience. In your code, after processing the test set, the frame ID and pedestrian ID are not preserved, only the (x,y) coordinates are used. Therefore the after sampling from the predicted distribution, we do not have the frame ID and pedestrian ID. Therefore, it is not very efficient to plot the trajectory. I am wondering do you have any suggestions to improve the efficiency of finding corresponding prediction results, frame ID and pedestrian ID. Thanks a lot! Bests, Xingchen 嗨 Abduallah @abduallahmohamed <https://github.com/abduallahmohamed>， 非常感谢您的回复。关于我在训练集和测试集中的负值，一些坐标是负的。但现在我知道没关系，我已经成功地绘制了轨迹。非常感谢您的回答，非常有用。 最后想问一下，不知道各位能否根据自己的经验给点建议。在您的代码中，处理完测试集后，不保留框架 ID 和行人 ID，仅使用 (x,y) 坐标。因此，从预测分布中采样后，我们没有帧 ID 和行人 ID。因此，绘制轨迹不是很有效。我想知道您有什么建议可以提高查找相应预测结果、帧 ID 和行人 ID 的效率。 非常感谢！ 最好的， 星辰 Hello， I've been learning about trajectory prediction recently, and I want this project to predict the trajectory of my own video data. Can you share your visualization code with me? If you can, please email me: ***@***.*** , be deeply grateful！ — Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFIVLG4NQ4ZDCPAD7RQTNXDVG2LNJANCNFSM4LCGS4JA> . You are receiving this because you were mentioned.Message ID: ***@***.***>