Assalam o Alaikum! excellent video. would you help me in fake videos detection and extracting frames. i would pay for this. please. i am not getting fake videos dataset. here is a source. can you provide dataset related fake videos.
That is just the video display error caused because the video size is bigger than the output limit set in the colab notebook, you do not have to solve it to run the application, you can just save/download the output video in your system and play it in any video player you like.
Excellent explanation of human activity recognition using deep learning, one of the best videos I've seen on the subject! You had shown pose estimation in the lead up to ConvLSTM and LRCM but did not use it. in your two examples Can the frame sequence of pose vectors (comprised of skeletal joint locations) be used with an LSTM to improve the accuracy even further. Although, there are some recent papers on using Graph Convolutional Networks for human activity recognition. How does the accuracy of GCN-based human activity recognition compare with LRCN?
Thank you Vijay for the complements, For the first part of your question, yes, you can try sending the skeletal points to further improvement, although I"m not sure if the increase in computation will be worth the accuracy gains, some people opt for optical flow and then feed the motion vectors to the model to capture motion information. For the second part, I've not done such a comparison so can't say about that.
@@alimchogle3876 apparently there are still ufc50 datasets from kaggle u can find. and I believe u can use them. I downloaded just to see how the structure is, so that I can properly organize my data directory
Very well explained video, I have searched this topic a lot, you really did it. Thank you very much. I have a question, Can we apply this to Real-Time Scenarios such as detecting and classifying plant diseases?
Having unexpected trouble from ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type generator) from my video dataset, do you have any suffestion?
Thank you sir for your wonderful explanation. I recently download the source code and also fill my email, but i didn't see the source code on my inbox. Could you please send how to download your code. I really appreciate for your kindness. Thank you.
thank you for your excellent videos, but the font size is small and hard to read could you please zoom In a lot more? also, the voice is not good, could you please use a microphone and also speak more slowly and clearly?
Dear Mr. Taha, I thank you from my heart to create such a great content. Your open source well-documented code and detailed walkthrough video of the implementation made me learn a lot of practical stuffs. I am working on a project that uses ConvLSTM to detect emotion from frontal face video and I found your video + code extremely helpful for my project, as I can carry forward the things I have learned from your video. The world needs more people like you! Please keep on making these wonderful videos and open source code. You don't know how much you are helping people in the field of Deep learning and Computer Vision! ❤
Dear Efran, I came across your comment while reviewing the discussion thread. I am currently working on a similar project, specifically developing a Video Facial Emotion Recognition model using ConvLSTM. I was wondering if you could share your findings with me. Thank you! Best regards, Sebastian
Hi sir appreciate if you could share your insights on my project. I am having 4 scenarios/action for recognition, and would like to develop a video analytics system (real-time) to ensure all 4 actions are executed with correct order. Will this CNN+LSTM method be suitable for this task? Thanks in advanced.
hello, I really like your explanation, and I want to ask you a favor. Can you please make a video about the MoViNet model for video classification? I would really appreciate it I am new to AI and really struggling :(
Hi, the link to the paper for using ConvLSTM layer is in the notebook, please download the code or go over the blogpost to get the exact link. Also the paper explains the layer, the architecture of the model itself in the tutorial is custom.
I am trying to use this code on a smaller dataset but it doesn’t provide high accuracy at the end and there is overfitting. What shall I do to prevent this?
Hi..At first ,,thanks for such a powerful video.. I have tried to make it real time but I m facing some errors.. ... Can You plzzz implement real time testing in this video . Plz plz.. I have seen so many TH-cam videos bt I m getting those videos ..I will be very grateful if you make it real time
Thank you Samia, at the moment I'm not going to publish a tutorial for real-time use case but I've already published an action recognition pipeline for realtime here: learnopencv.com/introduction-to-video-classification-and-human-activity-recognition/
This dataset given is not running, I tried extracting it and also tried entering the folder path but it is not running. Can anyone help to run this dataset?
Complete and patient training. 👍 Thank you very much for this very useful content.😊 I just verified the email, but I could not receive the codes. Can you please guide me?🙏
The world need more people like you. Love your practical approach to problems. Only suggestion is a external microphone. The voice will be great will a little more richness.
Hi, thank you for the detailed video and the implementation. Did you try the model with optical flow field input? As per the paper on LRCN, they are using a weighted output of two networks with RGB and optical flow field input which resulted in a much better performance. The LRCN RGB model is better in the classification of activities involving objects in the frame (e.g., typing-keyboard object). On the other hand, the flow field model is much better at classifying activities involving a lot of motion (e.g., Soccer Juggling, which includes more generic objects in the frame, but identifies based on class-specific motion cues).
No I have'nt but you're right, high motion based activity recognition will definitely benefit temporal motion information using something like optical flow or even SlowFast style networks
Hii... Final year student cs student here and thinking of doing violence detection from surveillance systems.. any information related to this like papers you find useful or videos or anything you know will be really helpful 🙏🙏
Hello sir, thank you for the good content, but i have tried to download the code it doesn't work. {Sorry! Your confirmation url is not valid} can you help me to get the link? Thank you
@@BleedAIAcademy thank you very much. And one more thing, can i use your code for my project? If that's okay with you, I don't want to get any copyright infringements.
hi i have tried downloading the video using youtube-dl but it shows an error in the uploader id as youtube has removed youtube-dl is there any alternate way to download the youtube video and saving it in mp4 format??? anyone?
hello,the Download Code is Real-Time 3D Hands Landmarks Detection & Hands Classification,not Human Activity Recognition using CNN + LSTM maybe something went wrong
@@BleedAIAcademy the link you have provided to download the code is wrong it is for Real time 3D Hands landmarks detection & classification not for the HAR
HI , Dear Taha ,is there a possibility to apply concept of single action detection and generalize it to detect more than one action ? Example video contains ( swing , running , sliding , jumping ) can we apply the same concept to your code and detect these movements in same video ?
Yeah, so there are dedicated approaches for that, but a hack to repurpose this approach for what you're trying to achieve would be something like this: Use people detection to get bounding boxes of all the people, then adjust the boxes for input size you need for the model, pass in the all the boxes to the model to get predictions. Now obviously this would be slow but to make it a bit fast, use an object tracker to detect activity every nth second or frame.
it took me 4 hours to understand taking video frames and storing it in list .. after that i tried by myself. it worked.. so just try to get concept thereotically first and then try to do it by urself.. just dont look code to copy. take pen and paper and write step by step what u need to do.. for eg. if u want to extract frames then write like this.. : to get into main dir i need to set path variable. then if i have 10 class then i need to loop through it. to get into first class . take its index. then list all dir..then select video , then extract.. like this .. to understand it properly .. just take one class and only 1 video first and outline what to do.btw. thanks sir for the turorial
Really well explained video, I have search alot for this topic you really nailed it. Thank you so much. I have a question , Can we implement this on a Real-Time Scenario like detecting a specific or multiple actions that we wanted to detect?
Thank you for your comments and Yes, Please do read the associated blogpost and I've written another one where I do this for a video too. But if you're looking to monitor CCTV footage and actions in real time of multiple people then this can quickly get complex, because when you're adding a temporal domain to the model the latency will take a hit. So do check pros and cons of different approaches out there to figure what's best for your particular use case.
hello sir thanks for your time to explore the action recognition in videos,but i have one query if any action not existing in dataset but am testing that activity its not give proper resut how to solve this problem sir
I found error at Construct the required LRCN model here: LRCN_model = create_LRCN_model() it shows NotImplementedError:Cannot convert a symbolic Tensor(lstm_5/strided_slice:0) to a numpy array.This error may indicate that you're trying to pass a Tensor to a Numpy call, which is not supported.
What kind of adjustement can be made to optimize the following to having multiple individuals in a scene and each getting detected and classified by the action they are doing.
A basic technique is to run person detection on the video, and then inside each person crop do action recognition. Optimize this by doing action recognition only after set intervals, since no one will be changing their actions every next frame, also track people instead of detecting on each frame. There are dedicated models for this specific problem in paperswithcode, this problem Is called: Spatio Temporal Detection, here's a nice dataset for it: arxiv.org/abs/1705.08421
@@BleedAIAcademy Thank you for the reference I will make sure to look it up. Will you be covering something similar anytime soon? For example action recognition in Sports.
Hi Thank you for the amazing content.but i have encountered these errors：1、NameError: name 'plot_metric' is not defined. 2、KeyError: 'dislike_count'，and your command did't work. what can i do to solve these errors?? extremely grateful
@@it_08amanagarwal35 uninstall Pafy library and then Install Pafy again with this command: pip install git+github.com/TahaAnwar/pafy.git#egg=pafy Make sure to restart the notebook/kernel.
super explanation. sir want to detect the play going on clips such passing the ball and running with ball on the field with no. of players are playing can i use this approach.
Dear Taha, I really want to thank you for creating such content and at no cost, so thank you so much. You have not only explained in a magnificent way, but also taught me how to write code in a way that eases understanding. I have downloaded your code and applied it ! do you think it possible to take it a step further and break down which videos did the model/s find hardest to classify and possibly use a grad-cam approach to see which activations were used to make predictions on successful predictions versus failures?? I would also be very interested to see how a 3DCNN would compare to these models (I know you mentioned in your blogpost that it's computationally expensive and slow but would there be an improvement in performance??).. Many thanks, D
Thanks D Ahmed for your kind words, And yes you can use CNN analysis methods like GradCAM and other to judge hard to classifer examples and retrain on them for better results, as for a 3DCNN I can't say for the performance compared to this plus it also depends upon which architecture / paper of 3dCNN you're trying to implement. Here's a great tutorial from pyimagesearch that leverages 3d CNN: www.pyimagesearch.com/2019/11/25/human-activity-recognition-with-opencv-and-deep-learning/ this will be useful for you.
i followed the instruction to download the code but could not get, even used 2 different email id and browser. but still did not get anything after verification of email. can u help me how can i get code.
Hi Sir, OSError: MoviePy error: the file test_videos/Comparison of Four Styles of Tai Chi-Output-SeqLen20.mp4 could not be found! how to solve this error!!
Dear Taha, thank you for giving such a great explanation. However, i have a question. is it possible for me to use the combined algorithm to predict (lets say) a growth of a tree by feeding it images from one week to another
Dear Mr. Taha, I am really appreciated of your video. I want learn about that and I need to download the code, but it doesn't send to my email. Could you help to send this code? thank you very much
This video helps me a lot! Thank you so so much....
when is the site going to start working again
Dear sir can we use the data as violence detection key Framing
copy paste explanation....!! does not makes sense what are you explaining or speaking only..
if i want to do it with cars and not people, what would you recomend guys, behavior of cars while driving
Can anyone download the code?
drive.google.com/drive/folders/1zt3Kt_KogaewsPBIbUPONQCScQWyBVKF?usp=drive_link
@@BleedAIAcademy 👌
amazing video , very detailed and organized .Can you use the same model to detect real -time using camera instead of uploading videos from youtube ?
hello can i use this on real time cctv video? for detecting Recognition of ppls on real time?
is it possible to do in live footage say for example live football match
Can this model be applied to the ucf-crime datset
permision to learn sir. how i can downlaod your dataset?
thanks before
I tried it but my accuracy was around 77 and it was giving wrong result and was flickering a lot
The link site does not send the source code
Assalam o Alaikum! excellent video. would you help me in fake videos detection and extracting frames. i would pay for this. please. i am not getting fake videos dataset. here is a source. can you provide dataset related fake videos.
hi, Taha. I want source code .
Hello the blogpost link is not working properly. Can you please check the link? It says the page not found
Thanks for pointing this out. Please check the link now.
@@BleedAIAcademy Thanks a lot
Iam not able to download the test video from the code u provided guved me regexnit found error
This explained video is very well, but I can’t understand how to solve the ValueError : max_duration at the end of the video ?
That is just the video display error caused because the video size is bigger than the output limit set in the colab notebook, you do not have to solve it to run the application, you can just save/download the output video in your system and play it in any video player you like.
Excellent explanation of human activity recognition using deep learning, one of the best videos I've seen on the subject! You had shown pose estimation in the lead up to ConvLSTM and LRCM but did not use it. in your two examples Can the frame sequence of pose vectors (comprised of skeletal joint locations) be used with an LSTM to improve the accuracy even further. Although, there are some recent papers on using Graph Convolutional Networks for human activity recognition. How does the accuracy of GCN-based human activity recognition compare with LRCN?
Thank you Vijay for the complements, For the first part of your question, yes, you can try sending the skeletal points to further improvement, although I"m not sure if the increase in computation will be worth the accuracy gains, some people opt for optical flow and then feed the motion vectors to the model to capture motion information. For the second part, I've not done such a comparison so can't say about that.
how do i connect the database please help?
Sadly the dataset is no longer available. Hope you can check this again, because I think this video may be of help still. (I will be watching this!)
Yes the dataset is no longer available and I am facing problem for that
@@alimchogle3876 apparently there are still ufc50 datasets from kaggle u can find. and I believe u can use them. I downloaded just to see how the structure is, so that I can properly organize my data directory
Very well explained video, I have searched this topic a lot, you really did it. Thank you very much. I have a question, Can we apply this to Real-Time Scenarios such as detecting and classifying plant diseases?
Thank you, and You don't need a video classification solution for this sort of task.
Having unexpected trouble from ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type generator) from my video dataset, do you have any suffestion?
Hi james, as mentioned in the error message, the data type you are using is not supported.
Can i do this with realtime video
Can you draw the confusion matrix for the 4 classifications. Can anyone help with the code
You are a great teacher. I selected this video very accidentally. Everything was clearly explained. Thank you very much.
Thanks
Cannot download the code.
After filling the email, The link is broken
Eran
Thank you sir for your wonderful explanation. I recently download the source code and also fill my email, but i didn't see the source code on my inbox. Could you please send how to download your code. I really appreciate for your kindness. Thank you.
Download code Link: drive.google.com/drive/folders/1zt3Kt_KogaewsPBIbUPONQCScQWyBVKF?usp=sharing
@@BleedAIAcademy Thank you so much.😄
thank you for your excellent videos,
but the font size is small and hard to read could you please zoom In a lot more?
also, the voice is not good, could you please use a microphone and also speak more slowly and clearly?
Noted!
Nice explanation, can you share your code? Thanks a lot
Code Download Link: drive.google.com/drive/folders/1zt3Kt_KogaewsPBIbUPONQCScQWyBVKF?usp=sharing
Dear Mr. Taha, I thank you from my heart to create such a great content. Your open source well-documented code and detailed walkthrough video of the implementation made me learn a lot of practical stuffs. I am working on a project that uses ConvLSTM to detect emotion from frontal face video and I found your video + code extremely helpful for my project, as I can carry forward the things I have learned from your video. The world needs more people like you! Please keep on making these wonderful videos and open source code. You don't know how much you are helping people in the field of Deep learning and Computer Vision! ❤
But I didn't get email for code after clicking on given link and varifying the email-id as mentioned in description. Can you help me pls
Thank you Erfan, I'm humbled by your praise 💖
Dear Efran,
I came across your comment while reviewing the discussion thread. I am currently working on a similar project, specifically developing a Video Facial Emotion Recognition model using ConvLSTM. I was wondering if you could share your findings with me.
Thank you!
Best regards, Sebastian
thanks!
Hi sir appreciate if you could share your insights on my project.
I am having 4 scenarios/action for recognition, and would like to develop a video analytics system (real-time) to ensure all 4 actions are executed with correct order. Will this CNN+LSTM method be suitable for this task? Thanks in advanced.
Thank you brother, this is the most comprehensive video on this subject out here!
Thank you
hello, I really like your explanation, and I want to ask you a favor. Can you please make a video about the MoViNet model for video classification? I would really appreciate it I am new to AI and really struggling :(
Bro that;s great...can i use these code for my research purpose??
Sure, But do give proper credit.
Brilliant Job, could,nt find code in the link. please check again
Can u tell me which paper did u follow to implement ConvLSTM model?
Hi, the link to the paper for using ConvLSTM layer is in the notebook, please download the code or go over the blogpost to get the exact link. Also the paper explains the layer, the architecture of the model itself in the tutorial is custom.
@@BleedAIAcademy can u tell me what learning rate are you considering?
Code is not working. Lots of errors during execution of the last 4 sections.
Can u send updated code.
I am trying to use this code on a smaller dataset but it doesn’t provide high accuracy at the end and there is overfitting. What shall I do to prevent this?
Hi..At first ,,thanks for such a powerful video..
I have tried to make it real time but I m facing some errors.. ...
Can You plzzz implement real time testing in this video . Plz plz..
I have seen so many TH-cam videos bt I m getting those videos ..I will be very grateful if you make it real time
Thank you Samia, at the moment I'm not going to publish a tutorial for real-time use case but I've already published an action recognition pipeline for realtime here: learnopencv.com/introduction-to-video-classification-and-human-activity-recognition/
I didn't get the email for the code after clicking on the given link and varifying the email-id as mentioned in the description.
This dataset given is not running, I tried extracting it and also tried entering the folder path but it is not running. Can anyone help to run this dataset?
Can u provide the goggle collab link and data set for this vedio sir?
Complete and patient training. 👍
Thank you very much for this very useful content.😊
I just verified the email, but I could not receive the codes. Can you please guide me?🙏
Please share your email address
I am unable to download the code by link (not receiving the email). Please help me.
The world need more people like you. Love your practical approach to problems. Only suggestion is a external microphone. The voice will be great will a little more richness.
Really helped sir. There are very few videos on video processing
Can you help with, human action prediction problems. The model will predict human action before it happens.
Hi, thank you for the detailed video and the implementation. Did you try the model with optical flow field input?
As per the paper on LRCN, they are using a weighted output of two networks with RGB and optical flow field input which resulted in a much better performance. The LRCN RGB model is better in the classification of activities involving objects in the frame (e.g., typing-keyboard object). On the other hand, the flow field model is much better at classifying activities involving a lot of motion (e.g., Soccer Juggling, which includes more generic objects in the frame, but identifies based on class-specific motion cues).
No I have'nt but you're right, high motion based activity recognition will definitely benefit temporal motion information using something like optical flow or even SlowFast style networks
Hii... Final year student cs student here and thinking of doing violence detection from surveillance systems.. any information related to this like papers you find useful or videos or anything you know will be really helpful 🙏🙏
I didnt get Source code Email, I think its not working!
where is method 1 ?
Hello sir, thank you for the good content, but i have tried to download the code it doesn't work. {Sorry! Your confirmation url is not valid} can you help me to get the link? Thank you
I'm not able to get the code? I tried three or four times.
Code Link: drive.google.com/drive/folders/1zt3Kt_KogaewsPBIbUPONQCScQWyBVKF?usp=sharing
@@BleedAIAcademy thank you very much. And one more thing, can i use your code for my project? If that's okay with you, I don't want to get any copyright infringements.
hi i have tried downloading the video using youtube-dl but it shows an error in the uploader id as youtube has removed youtube-dl is there any alternate way to download the youtube video and saving it in mp4 format??? anyone?
Did you find solution for it
Please help with this similar problem. Suggest any other library
hello, i am doing it in a jupyter note book in that i am getting video_path error can you explain y it is giving this error?
hello,the Download Code is Real-Time 3D Hands Landmarks Detection & Hands Classification,not Human Activity Recognition using CNN + LSTM
maybe something went wrong
It's working fine. What problem you're facing?
@@BleedAIAcademy the file I download from the above URL is your other project, not suitable for what this video explains
@@BleedAIAcademy the file I download from the above URL is "Real Time 3D Hands Landmarks Detection & Classification"
@@BleedAIAcademy the link you have provided to download the code is wrong it is for Real time 3D Hands landmarks detection & classification not for the HAR
@@sohaibshaikh5534 Oh, just checked by clicking the link, you're right.
hi i got an error ("x= features_train, y = labels_train, " ) are not define . code("convlstm_model_training_history = convlstm_model.fit(x = features_train, y = labels_train, epochs = 50, batch_size = 4,
shuffle = True, validation_split = 0.2,
callbacks = [early_stopping_callback])"
kindly resolve my error.
HI , Dear Taha ,is there a possibility to apply concept of single action detection and generalize it to detect more than one action ?
Example video contains ( swing , running , sliding , jumping ) can we apply the same concept to your code and detect these movements in same video ?
Yeah, so there are dedicated approaches for that, but a hack to repurpose this approach for what you're trying to achieve would be something like this:
Use people detection to get bounding boxes of all the people, then adjust the boxes for input size you need for the model, pass in the all the boxes to the model to get predictions.
Now obviously this would be slow but to make it a bit fast, use an object tracker to detect activity every nth second or frame.
my question is dont we need batches?
it took me 4 hours to understand taking video frames and storing it in list .. after that i tried by myself. it worked.. so just try to get concept thereotically first and then try to do it by urself.. just dont look code to copy. take pen and paper and write step by step what u need to do.. for eg. if u want to extract frames then write like this.. : to get into main dir i need to set path variable. then if i have 10 class then i need to loop through it. to get into first class . take its index. then list all dir..then select video , then extract.. like this .. to understand it properly .. just take one class and only 1 video first and outline what to do.btw. thanks sir for the turorial
How we can evaluate the model on the videos downloaded from youtype ?
Really well explained video, I have search alot for this topic you really nailed it. Thank you so much. I have a question , Can we implement this on a Real-Time Scenario like detecting a specific or multiple actions that we wanted to detect?
Thank you for your comments and Yes, Please do read the associated blogpost and I've written another one where I do this for a video too. But if you're looking to monitor CCTV footage and actions in real time of multiple people then this can quickly get complex, because when you're adding a temporal domain to the model the latency will take a hit. So do check pros and cons of different approaches out there to figure what's best for your particular use case.
Can you be a good man and also share the model saved after training please?
Can this be used in Lane departure warning system?
hello sir thanks for your time to explore the action recognition in videos,but i have one query if any action not existing in dataset but am testing that activity its not give proper resut how to solve this problem sir
Which algorithm is used here for CNN and LSTM
I am not able to get the source code. Can someone share it please
Thanks great explanation, do we have to have dropout.
where is the video file function?
it's not even in your code but used for output
Sir can you please provide the code
I found error at Construct the required LRCN model here: LRCN_model = create_LRCN_model() it shows NotImplementedError:Cannot convert a symbolic Tensor(lstm_5/strided_slice:0) to a numpy array.This error may indicate that you're trying to pass a Tensor to a Numpy call, which is not supported.
i want to create a novel model with help of frcnn and svm. can you guide me with it please?
Thank you sir for the beautiful explanation. Sir, can you help me in choosing the right approach for anomaly detection. Kindly make a video on it too
Sir how to take input from a live camera or a video we provide?
Taha u have done an amazing job..
What kind of adjustement can be made to optimize the following to having multiple individuals in a scene and each getting detected and classified by the action they are doing.
A basic technique is to run person detection on the video, and then inside each person crop do action recognition. Optimize this by doing action recognition only after set intervals, since no one will be changing their actions every next frame, also track people instead of detecting on each frame.
There are dedicated models for this specific problem in paperswithcode, this problem Is called: Spatio Temporal Detection, here's a nice dataset for it: arxiv.org/abs/1705.08421
@@BleedAIAcademy Thank you for the reference I will make sure to look it up. Will you be covering something similar anytime soon? For example action recognition in Sports.
@@rolandsalloum8797 No plans at the moment.
Hi Thank you for the amazing content.but i have encountered these errors：1、NameError: name 'plot_metric' is not defined. 2、KeyError: 'dislike_count'，and your command did't work. what can i do to solve these errors?? extremely grateful
I'm facing same issue.
Please tell me the solution asap!!
@@it_08amanagarwal35 sorry to hear that,i don't solve these error yet.
@@it_08amanagarwal35 uninstall Pafy library and then Install Pafy again with this command:
pip install git+github.com/TahaAnwar/pafy.git#egg=pafy
Make sure to restart the notebook/kernel.
@@jackiechen2313 check the comment
@@jackiechen2313 have you got solution?
super explanation. sir want to detect the play going on clips such passing the ball and running with ball on the field with no. of players are playing can i use this approach.
Why after applying first convolution , out put is 64 ? It should be 62*62*16 right
Sir I am getting error in download youtude-dl videos it is showing unable to extarct uploader id how to solve this issue
did you have a solution? i have same problem too
This video should be famous!
Thank you Kunal :)
@@BleedAIAcademy Hello sir! Do you have any video in which I can use AGG16 with LSTM? For video classification?
@@kunaltulsidasani6766 Nope
@@kunaltulsidasani6766 Not exactly
@@BleedAIAcademy what all architecture can I use which will take 4 inputs as shown here
Hello bro, can you please make a video on recurrent yolo on simple example any object tracking example??
Excellent video, thank you!
how can i use my pretrained weighst
Can't get the code after sharing / verification of email
your email address?
thank you very much sir, can we use this approach for fall detection ? please help me
great video.
how can i create a confusion matrix for this code
Dear Taha, I really want to thank you for creating such content and at no cost, so thank you so much. You have not only explained in a magnificent way, but also taught me how to write code in a way that eases understanding. I have downloaded your code and applied it ! do you think it possible to take it a step further and break down which videos did the model/s find hardest to classify and possibly use a grad-cam approach to see which activations were used to make predictions on successful predictions versus failures?? I would also be very interested to see how a 3DCNN would compare to these models (I know you mentioned in your blogpost that it's computationally expensive and slow but would there be an improvement in performance??)..
Many thanks,
D
Thanks D Ahmed for your kind words, And yes you can use CNN analysis methods like GradCAM and other to judge hard to classifer examples and retrain on them for better results, as for a 3DCNN I can't say for the performance compared to this plus it also depends upon which architecture / paper of 3dCNN you're trying to implement.
Here's a great tutorial from pyimagesearch that leverages 3d CNN: www.pyimagesearch.com/2019/11/25/human-activity-recognition-with-opencv-and-deep-learning/ this will be useful for you.
hiw can i use my pretrained weights
i followed the instruction to download the code but could not get, even used 2 different email id and browser. but still did not get anything after verification of email. can u help me how can i get code.
Hi Sir,
OSError: MoviePy error: the file test_videos/Comparison of Four Styles of Tai Chi-Output-SeqLen20.mp4 could not be found!
how to solve this error!!
Hi, make sure this exact file is present in the directory specified, try renaming the file and double check the path.
HI , can u help me in understanding the dataset please
Dear Taha, thank you for giving such a great explanation. However, i have a question. is it possible for me to use the combined algorithm to predict (lets say) a growth of a tree by feeding it images from one week to another
This video is really cool!, I could understand ConvLSTM2D very clearly, I subscribed your channel :)
can we do activity recognisation using YOLO and CNN-LSTM
Yes, if you want to focus on specific regions of the videos to detect activities.
hi u explained very well can u just let me know how can i use this for UCI HAR datset
Can you add auto subtitles for your videos?
Great work man! keep sharing with us
Thank you Bukhari Jee
Very well illustrated. Thanks for putting this together.
Thank you
Dear Mr. Taha, I am really appreciated of your video. I want learn about that and I need to download the code, but it doesn't send to my email. Could you help to send this code? thank you very much
Please share your email.