I want to use this demo to reappear the results on ETHZ dataset but the results aren't good

[Sunny-Yu]

Hello, I want to use this demo to reappear the results on ETHZ dataset. The results of the demo data are OK, but when I use the demo code on ETHZ, the results aren't good enough. The most prominent point is there are several bounding boxes on the same people at on frame. It seems that the results lack of nms treatment. I use the rectangles120x300.pom from https://github.com/pierrebaque/generatePOMfile and also use the rectangles480x1440.pom from ETHZ dataset, but both of them can't reappear the results on the paper. I think maybe some parameters should be modify from demo config. I hope you can help me which parameters should be noted. Thank you!

[GehenHe]

This also disturb me a lot

[pierrebaque]

Hello,

Thank you for your message. What are the parameters that you don't manage to set? What kind of results do you get? Does it look like the rectangle indices are not correct somehow or the tracking area is not good?

I also used rectangles120x300.pom. I put the file back directly in the Git here.

In RunPOM.ipynb, you should use something like
a,p,alpha_black = 1.5,250,0.8

Did you check the background/foreground and parts segmentation? Do they look reasonable?

You can send me some detection images if you want, maybe I can try to guess what is wrong. Did you retrain everything?

@gooa1121 do you have an idea of what may have changed?

If it really does not work, I will send you directly my trained models.

Best,
Pierre

[Sunny-Yu]

Hello,

Thank you for your reply.

Firstly, about the parameters, I want to know if the parameters of Config.py need to be modified when I detect ETHZ dataset. For example, H_grid and W_grid should be changed to 120 and 300. The Config.py is from the tozip file. I want to know which parameters are also to be modified.

Secondly, about the result from RunPOM.ipynb, I realize that I mistakenly used the foreground results as detect results after reading your reply. I attached the foreground result to the following which is generated by the code 'room.plot_output(Q_out,0,0,8,thresh = 0.95,iteration=-1)', and a,p,alpha_black = 1.5,200,0.5.

So, another question, which part's result is the detect result? The Run POM? Or the Run Baseline?

Thank you very much for helping me. You are so kind.

Best,
Sunnyu

[gooa1121]

Hi,
About Config file, if you're not retraining any new model, it should be fine to keep other parameters unchanged except POM size and POM file path.
Trying different combination of values for parameters a, p, alpha_black might help. You can also change the iteration parameter to other values in room.plot_output to check if the body part segmentation works well.

Best,
Pei-I

[wanmok]

Hi,

If I want to run Wildtrack_dataset, do I have to change something? I encountered some strange dimension mismatch problem. It says:

Apply node that caused the error: GpuDownsampleFactorMax{(2, 2),True}(GpuElemwise{Composite{maximum(i0, (i1 + i2))}}[(0, 1)].0)
Toposort index: 352
Inputs types: [CudaNdarrayType(float32, 4D)]
Inputs shapes: [(1, 64, 1072, 1920)]
Inputs strides: [(0, 2064228, 1922, 1)]
Inputs values: ['not shown']
Outputs clients: [[GpuDownsampleFactorMax{(2, 2),True}(GpuDownsampleFactorMax{(2, 2),True}.0), GpuContiguous(GpuDownsampleFactorMax{(2, 2),True}.0)]]

HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'.
HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.

Could you give any clue to fix this? Thanks!

[TTTREE]

Hi,

If I want to run Wildtrack_dataset, do I have to change something? I encountered some strange dimension mismatch problem. It says:

Apply node that caused the error: GpuDownsampleFactorMax{(2, 2),True}(GpuElemwise{Composite{maximum(i0, (i1 + i2))}}[(0, 1)].0)
Toposort index: 352
Inputs types: [CudaNdarrayType(float32, 4D)]
Inputs shapes: [(1, 64, 1072, 1920)]
Inputs strides: [(0, 2064228, 1922, 1)]
Inputs values: ['not shown']
Outputs clients: [[GpuDownsampleFactorMax{(2, 2),True}(GpuDownsampleFactorMax{(2, 2),True}.0), GpuContiguous(GpuDownsampleFactorMax{(2, 2),True}.0)]]

HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'.
HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.

Could you give any clue to fix this? Thanks!
I am trying to do the same thing about running Wildtrack_dataset. Have you succeeded?

[Sunny-Yu]

Hi, I didn't get a good result in the end, maybe it was not set up correctly in some places. And  time has passed a bit, most of the details have been forgotten, so I'm sorry that I can't help you. Hope the original author can help you.

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"TTTREE"<[email protected]&gt;; 发送时间:&nbsp;2020年3月25日(星期三) 中午1:54 收件人:&nbsp;"pierrebaque/DeepOcclusion"<[email protected]&gt;; 抄送:&nbsp;"844758999"<[email protected]&gt;;"Author"<[email protected]&gt;; 主题:&nbsp;Re: [pierrebaque/DeepOcclusion] I want to use this demo to reappear the results on ETHZ dataset but the results aren't good (#31) Hi, If I want to run Wildtrack_dataset, do I have to change something? I encountered some strange dimension mismatch problem. It says: Apply node that caused the error: GpuDownsampleFactorMax{(2, 2),True}(GpuElemwise{Composite{maximum(i0, (i1 + i2))}}[(0, 1)].0) Toposort index: 352 Inputs types: [CudaNdarrayType(float32, 4D)] Inputs shapes: [(1, 64, 1072, 1920)] Inputs strides: [(0, 2064228, 1922, 1)] Inputs values: ['not shown'] Outputs clients: [[GpuDownsampleFactorMax{(2, 2),True}(GpuDownsampleFactorMax{(2, 2),True}.0), GpuContiguous(GpuDownsampleFactorMax{(2, 2),True}.0)]] HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'. HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node. Could you give any clue to fix this? Thanks! I am trying to do the same thing about running Wildtrack_dataset. Have you succeeded? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[leconghieu99]

@Sunny-Yu can you please send me the .zip files. Looks like the link is broken. Email: [email protected]

[Sunny-Yu]

Hi, I don't save the files, please request from the owner.&nbsp;

…

---Original--- From: ***@***.***&gt; Date: Wed, Feb 9, 2022 02:52 AM To: ***@***.***&gt;; Cc: ***@***.***&gt;;"Xufeng ***@***.***&gt;; Subject: Re: [pierrebaque/DeepOcclusion] I want to use this demo to reappear the results on ETHZ dataset but the results aren't good (#31) @Sunny-Yu can you please send me the .zip files. Looks like the link is broken. Email: ***@***.*** — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.Message ID: ***@***.***&gt;