hbilen / dynamic-image-nets Goto Github PK

from the paper we can see that the model can only give the average of all the classes and evaluate the whole model , so how I can get the single class such as Biking Class accuracy ? thanks for your help !

how can i see the dynamic image result? after run the step 5

How did you extract optical flow and DOF

I am a newcomer to programming. After the fifth step, I want to compute approximate dynamic images and get the dynamic image, but I didn't find the relevant code. I don't know what to do next. Thank you for your help.

Hi, thanks for the code.
Let say, I have one video from UCF101, then I extract the optical flow images from that video. Can I use trained model (Deploy\resnext50-of-arpool-split1.mat) and run cnn_dicnn_of() to predict predifined activity within this video? Thank you

Hi, I find a strange problem when evaluating your code on ucf101. The accuracies are shown below.
Accuracy on split 1: 0.659001
Accuracy on split 2: 0.996249
Accuracy on split 3: 0.997294

Do you have any idea about the huge difference?

Love the work, I am just having difficulty understanding the architecture for the SI + DI model.
From what I see in the architecture of the resnext.mat model, the model uses a temporal max pooling layer just before the softmax layer. It says the input to the temporal max pooling layer are the merged conv7 features and Video2. I am assuming the merged conv7 features come from running the dynamic image through the ResNext model. Where does the Video2 come from?
Are we supposed to pass the whole video or just a single frame from the video clip?

i want see the dynamic image ,but in the code" vl_nnarpooltemporal layer"get the image is not same with the paper's image.the left image is from paper, the right image is from matlab code.can you help me?

How can I reproduce the results in the paper?
I could train the dynamic image net.
However, I could not find prediction or evaluation script in the code.
Could you please tell me how to do the evaluation? Or can you upload the evaluation code?

function di = compute_approximate_dynamic_images(images)
% Computes approximate dynamic images for a given array of images
% IMAGES must be a tensor of H x W x D x N dimensionality or
% cell of image names

I want to use the approximate_dynamic_images. I do not know how to choose N for a video.

I don't understand the SI(RGB image)
what the different between the DI: dynamic RBG image and SI: RGB image?

Hello, I am very interested in your work. I got a problem that the link you provide to download the cnn models for the ucf101 dataset is invalid, can you provide the link again? Thank you very much!

Hi, I have two questions to ask you，thanks

1. How to extract the optical flow diagram
2. How to combine SI+DI+OF+DOF to get 95% accuracy

please show some sample file list in README.md like below

data/UCF101/ucfTrainTestlist/
├── classIndFixed.txt
├── classInd.txt
├── testlist01.txt
├── testlist02.txt
├── testlist03.txt
├── trainlist01.txt
├── trainlist02.txt
└── trainlist03.txt

Convert videos to frames, resize them to 256x256 and store them in such a directory structure

is it possible to train on CPU.?

hey, recently i am reading your paper about dynamic image net, and one question i do not know how to achieve it:
it seems that the dynamic images are obtained offline, and i do not how to obtain this.

Hello,
I wanted to re-implement the function compute_approximate_dynamic_images with python and compare the results. But I can't find the function visualize_approximate_dynamic_images in the matlab code...
I tried to perform L2 normalization on the output of compute_approximate_dynamic_images, and multiply the values by 255 to get an image. Is this the correct way of visualizing dynamic image?

while I run the cnn_dicnn， I get the following error，but I have no idea;

cnn_dicnn
train: epoch 01: 1/1193:Cell contents reference from a non-cell array object.

Error in cnn_video_get_batch (line 23)
fetch = numel(images) >= 1 && ischar(images{1}) ;

Error in cnn_dicnn>getDagNNBatch (line 177)
im = cnn_video_get_batch(images, VideoId1, opts, ...

Error in cnn_dicnn>@(x,y)getDagNNBatch(bopts,useGpu,x,y) (line 102)
fn = @(x,y) getDagNNBatch(bopts,useGpu,x,y) ;

Error in cnn_train_dag>processEpoch (line 201)
inputs = params.getBatch(params.imdb, batch) ;

Error in cnn_train_dag (line 87)
[net, state] = processEpoch(net, state, params, 'train') ;

Error in cnn_dicnn (line 73)
[net, info] = cnn_train_dag(net, imdb, getBatchFn(opts, net.meta), ...

thank you for your help!

hello，i want to know how the accuracy is computed when the input modility is MDI? For SDI, a dynamic have one recognition result while MDI have more than one, in the paper, it seems not to be introduced how
the MDI accuracy is made. Thanks for your reply.

I am trying to train own data using your model. My dataset contains 29398 folders, each of which has 11 frames. There are two classes: normal and abnormal. I wrote the custom sepup_data.m for it. When I run main_train.m, some errors occurred as follow. How can I solve it? Will I need to change the parameters of architecture to fit this dataset? Thanks for your time!

How can I use other DL frameworks to achieve this, such as tensorflow？

Is it possible?

I have visualize the dagnn by this. But I can not re-implement the last few layers such as alPooling, ultiClass, kPooling and so on.

Do you have any idea about this?