pierrebaque / deepocclusion Goto Github PK

I'm trying to run example notebooks, but provided data and Config.py are not compatible. Config.py references data that are neither in repository nor in provided data archive.

In the RunUnaries notebook you pointed to a dataset:
im = Image.open('/cvlabdata1/cvlab/datasets_people_tracking/ETH/day_2/annotation_final/cvlab_camera2/begin/00000100.png')

where can i download this dataset?

Hi, thank for your great work!
it seem that i run this project without error, but it blocks on the RunPom file(multithread model).
Am I wrong? And what is the result of this project
thanks

I tried to run first notebook RunParts.ipynb.
After I figured out that I need to use theano 0.9 and python 2.7 I've got stuck on missing import Config.
I tried to create it on my own, but I don't know what it's content should be.

I found that there are 6400 rectangles for each cam in the file rectangles80x80.pom.

and I have read the code to produce .pom file from：
https://github.com/pierrebaque/generatePOMfile

If I want to get the pom file, do I need to place a previous chessboard in ground plane to obtain the camera calibration files (OpenCV format)

Thanks for your time

Thank for your great work!

about your work, can I ask some notation?

In 3.2.1 simple generative model,
equation(2) ... if Z_i = 0 all i in L_k^c
I think that L_k^c has Ti and Bi, so Zi is always 1...
why is there Z_i=0 in L_k^c?

what did I misunderstand..?

Is Z in RoIpooling's parameter? like objectness?... but it seems not..

Thanks, It is great work. What's the python, Theano version required?

Hello!
Thank you for your work!
I have a question about algo perfomance that I was not able to find in the paper, what are the numbers in general, what are the heaviest parts in the online step, and how does it depend on GPU perfomance?

When i run that code step by step i am facing this issue again #15
But i am having GPU in my system. Can anyone help me to solve this issue..

How can I obtain my own pickle files in VGG/models base on my dataset ?

This code was developed before 2 years. Within these two years a lot version update for each package happened. When i am trying to install a lot of mismatch happened in each package.
Can anyone help me out solve this issue by providing dependencies with corresponding Package version.

When i run RunUnaries.ipynb, this error occurred. So i find some answer in google. Someone said that upgrade the Theano , and it did work, but another error occur and the terminal suggest us to browse https://github.com/Theano/Theano/wiki/Converting-to-the-new-gpu-back-end%28gpuarray%29. It says something about cuda, but mac needs cuda? it trouble me a lot. How can i solve this problem —— ImportError: cannot import name as_cuda_ndarray_variable.
Thanks

Hi pierrebaque,

I am confused with the 'parts' in the RunParts.ipynb, what does the 'parts' mean? and what does the RunParts.ipynb do? Why there are 9 parts? Thank you for your time.

when I run Runparts.ipynb , this error occurred. I think "Config" is a configuration file , but I can't find this file
could you please help me ?
thanks !

Hi @pierrebaque,
Thank you for your help. I set up all the configs and environment following README. The result of 'run_pom' function is Q_out, whose element is 1 or 0. I wonder whether the result I got is correct or not. I mean I thought the result would be probability, float number like 0.9, 0.8, 0.1 but it is binary. So I just want to double check. Thank you.

Hi, thanks for sharing your great work!

I downloaded a ZIP file from Google Drive, and was able to run the .ipynb codes successfully.
However, I am a bit confused.

The header section of "rectangles80x80.pom" is:

ROOM 80 80 4 1600

I suppose the "ROOM" header means
img_width = 80, img_height = 80, number of cameras = 4, number of locations = 1600 according to [1] and [2]

[1] https://cvlab.epfl.ch/software/pom
[2] https://github.com/pierrebaque/generatePOMfile

However, there seem to be 6400 locations as long as I looked at subsequent "RECTANGLES" sections of "rectangles80x80.pom"
Which is the correct number of locations, 6400(=80x80??) or 1600?

Plans to fix this in the future?

Hi!
This is a nice project!
I am trying to run this code on WildTrack Dataset to get the detection results. And I don't know if I need to modify the config.py or other files. So would you please tell me how to run this DeepOcclusion on WildTrack Dataset.
More detail better!
Thanks a lot !!!!!

When running the Unaries notebook, it has a dependency on what I believe is https://github.com/ddtm/theano-roi-pooling. It would be good to capture that in the readme.

There are also a number of imports referencing theano.tensor.signal.downsample in the codebase which is no longer available in current release 0.9 of theano. There's no code that actually uses the import, but it does stop the notebooks from running. I can make a pull request for those if needed.

I have run successfully the RunParts.py but something go wrong with the RunUnaries.py, which use the ROI-pooling layers. theano.sanbox.cuda was not used and it went wrong with:
from theano.sanbox.cuda import as_cuda_ndarray_variable, GpuOp
still error when I use:
from theano.gpuarray import as_cuda_ndarray_variable, GpuOp

how can I solve it ? Looking forward to your response!

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!

TypeError
in ()
1 import UnariesNet
----> 2 uNet = UnariesNet.unariesNet()

/home/kumar/DeepOcclusion/UnariesNet.py in init(self, load_pretrained)
80 c53_r = mNet.c53_r
81
---> 82 op = ROIPoolingOp(pooled_h=p_h, pooled_w=p_w, spatial_scale=1.0)
83
84

/home/kumar/DeepOcclusion/roi_pooling.py in init(self, pooled_h, pooled_w, spatial_scale)
18
19 def init(self, pooled_h, pooled_w, spatial_scale):
---> 20 super(ROIPoolingOp, self).init(self.func_file,
21 self.func_name)
22

TypeError: super(type, obj): obj must be an instance or subtype of type

How can i fix this ? Help will be highly appreciated ?

Could someone attach requirement.txt for the version used to emulate the project alsong with cuDNN and cuda version? That would be great as the author has *.txt in .gitignore

Running python 2.7 on ubuntu, running the first notebook RunParts.ipynb I get an error:

Running
parts_predictor.run_inference(bg_pretrained = True,regression_pretrained = True,verbose = True)

/DeepOcclusion/VGG/VGGNet.py

paramsValuesVGG = pickle.load(open(os.path.dirname(__file__)+"/models/paramsVGG.pickle","rb"))

Throws the error:
ValueError: insecure string pickle

Loading the pickle load on that file using instead my Python 3.6 environment gives:

>>> pickle.load(open("/home/ben/software/DeepOcclusion/VGG/models/paramsVGG.pickle","rb"))                                                                                                                                                                    
Traceback (most recent call last):                                                                                                                                                                                                                                             
  File "<stdin>", line 1, in <module>                                                                                                                                                                                                                                          
UnicodeDecodeError: 'ascii' codec can't decode byte 0xe2 in position 0: ordinal not in range(128)       

And if I change encoding to latin1/bytes gives:

pickle.load(open("/home/ben/software/DeepOcclusion/VGG/models/paramsVGG.pickle","rb"))
Traceback (most recent call last):
File "", line 1, in
UnicodeDecodeError: 'ascii' codec can't decode byte 0xe2 in position 0: ordinal not in range(128)

Which has already been reported and closed in #9 but I am unsure what the solution was.

Hi,
The BG output size is resized 4x4 times smaller than input according to VGG model(self.VGGout_resize = 16), resulting in a loss of detail.
Is there a way to get a BG with size not re-scaled?

Hi,
When I run the demo, I come across a problem,::::::
File ".\DeepOcclusion\DeepOcclusion\VGG\VGGNet.py", line 98, in init
paramsValuesVGG = pickle.load(open(os.path.dirname(file)+"/models/paramsVGG.pickle","rb"))
UnicodeDecodeError: 'ascii' codec can't decode byte 0xe2 in position 0: ordinal not in range(128)

It seems encoding is wrong, which should I use?

Thanks?

Can you give me the link of the driver you shared in this post?

As I run with ./RunParts.ipynb in the root directory
I got this error

./RunParts.ipynb: 1739: ./RunParts.ipynb: Syntax error: end of file unexpected (expecting "}")

What may cause this problem?

When I run the RunPom notebook, it expects to find a gaussian params file.

IOError: [Errno 2] No such file or directory: './Potentials/Parts/Run_transfer/gaussian_params.txt'

[theano v.0.9.0dev1.dev-2...]; changed

from theano.tensor.signal import downsample
...
downsample.max_pool_2d( ... )

to

from theano.tensor.signal import pool
...
pool.pool_2d( ... )

I am trying to run the code on a different dataset, RunParts works perfectly, however when I try to run unaries I get an error when rois_np in run_bulk_features is empty. The reason its empty is because in get_rois method, indices_reduced is empty as there are no values above the set threshold.

Is there something wrong with my setup to get this or is there a way to handle the exception when rois is empty?

IOError: [Errno 2] No such file or directory: './Potentials/Parts/Run_transfer/c0/2000.npy'
where is the file ?
thank you !

What version of theano, python, intel mkl is this? Could someone who has compiled and run this successfully do a pip freeze?

There is no Config module so this error keeps showing when trying to run.

Traceback (most recent call last): File "EM.py", line 36, in <module> from pom_room import POM_room File "/py/DeepOcclusion-master/pom_room.py", line 9, in <module> import Config

please help,i can't find it.

when I run RunParts.ipynb ,mistake 'TypeError:Expected an integer'
how to fix it ?

Hello,

Thank you for your work!

I have a quick question: I think the answer to the question is yes, but I wanted to confirm in case I set up something incorrectly:

Is it normal for the Deep Occlusion model to run really slow when it is used on CPU only?

Thank you!

Hi, when I run RunUnaries,
uNet.run_bulk_features(save_features = False)

It come to this:

IOError: [Errno 2] No such file or directory: './Potentials/Unaries/RunFeatures_transfer/00002000.npy'

How can I fix it? You response is greatly appreciated.

I want to run your code on the CPU successfully, but when I tried to run it on the GPU, a large amount of problems occurred.
I guess it's caused by the wrong version of Theano or CUDA drivers.
Could you tell me the version you used when coding the program?

Thanks

Hi! congratz for the great work.

Are you planning to share training code? I'm not able to reproduce results using current code because of old theano/cuda dependencies.

mistake in VGGNET.py line 39

Can you tell what is that mistake..

Originally posted by @DhanasekarChellamuthu in #34 (comment)

Hi!
When I'm running the RunPom part in the WILDTRACK dataset, the inference step become very sloooooow!
It's usually several minutes per frame!
I don't know what happens but it confuses me.
@gooa1121 @pierrebaque

Hi,

I have two questions about background model training:

1. I wonder whether I need to train from scratch or train from pretrained model? The pretrained model will make the convergence faster?
2. approximately how many images do we need to train my own model?