Currently, we have only processed facial recognition for the unmasked datasets and the frame rate is around 17-22fps. However, it is also dependent on the resolution of the webcam you are using.
10,000 images are decent enough to get a minimum level of accuracy however since we are performing facial recognition on two factors unmasked and masked with different points for landmark detection. Nearly 30,000 images in batches would be more ideal. Each batch is divided into 3 sets validation, training, and testing in which validation and training set keep changing with each forward and backward propagation so we can train it through more random collection. Thus cross-validation on batches of data containing 30,000 images was ideal for us.
In the poster, you mentioned splitting the data set into a training set and a validation set. Do you mean a training set and a testing set? There seem to be a lot of hyper-parameters to tune. Why is the split not the usual combination of a validate set, a training set, and a testing set?
Yes I agree by what you are saying and yes the split has been done into three sets. A validation set, a training set and a testing set which is the common split.
Thank you for spotting the error. We will change it in the poster.
YIP, Kam Wai
January 19, 2022 3:07 pm
Is there a pair of color and thermal cameras for each algorithm? Are the algorithms running simultaneously and being selected/aggregated to form the output?
Thanks for your question, yes there are two pairs of coloured and thermal cameras to produce one single output, and yes the algorithm will be an infinite functional loop. This means on every new person detected the algorithm will keep running itself to get the best results.
I do not understand the part that “algorithm to select camera depends on accuracy” and “The existence of
multiple cameras
allows multiple outputs
which will be compared on another algorithm and the most
accurate result will be displayed on a monitor”. Would you mind explaining them further? Will there be multiple algorithms running simultaneously in the final system?
Yes, there will be multiple algorithms running however but will be sequential so face recognition is the first followed by the mapping of the pair of thermal cameras and RGB colored cameras, and finally the selection algorithm. Here we look at the most accurate results on the facial landmark, however this is still in development phase so currently we don’t have the data to support this through empirical evidence. The idea is to select the most accurate temperature through facial landmark temperature mapping, our idea basically checks the set of facial landmark checkpoints and sees the one with the minimum difference from each other so we can select that pair of colored and thermal cameras.
We will work on that but due to covid restrictions, our entire team is not available to work on the hardware so we are doing multiple objectives not in proper order so currently uncorrelated tasks are being done not in sequence, therefore, I can understand your confusion. We will for sure have more coherent progress when we integrate all those sections. Thanks for your comments
KHURANA, Param
January 19, 2022 2:37 pm
Hello Mr. Singh,
Could you tell us some technical challenges your team faced in this project and how you overcame them. Thanks!
Hello Mr. Khurrana,
We have divided our work into multiple objectives. In hardware installation, the first is the driver required to be able to read data from the JTAG USB else an error regarding detection would be there. These are drivers for the different OS on ftdichip, this will allow the JTAG pod to be visible. Secondly, during the landmark facial detection the lack of our program to use masked individuals presented an obstacle for which we had to increase the dataset for our pre-trained machine learning model. Currently, we are facing issues regarding the mapping between the pixels of the coloured camera and thermal camera so we can accurately read the temperature of the facial landmarks featured detected during the facial recognition step.
I saw that there is an objective statment ” To identify the Optimum angle placements of the Camera to maximize the range and enhance the detected results”
Will the algorithm calculate the optimum angle placement? Or this will be calculated/found by the person who setup the system?
Also, may I have more detail on the “landmark features” such as images of them?
I guess I saw similar images on other group that also supervised by Professor Albert Wong. But it seems it was not shown on your poster.
Thanks for your question. The configuration will be manual however we will check the accuracy through the detection of landmark features from the AI algorithm and the setting with the most accurate temperature will be selected.
We did not mention the landmark features in our poster as we did not reach this part when we made the poster. However, now we have completed it and if you would like to see it, please send me your email details if possible as I cannot attach images on this comment.
Thanks for both of your reply, I guess you can use services such as imgur to post the link of your image files or even new version of poster.
So that other viewers can also take a look on your updated results.
Thanks for your question. For the angle, it will be set manually but there will be an algorithm that would check if all the landmark points are being identified and can detect fever. Coming to the landmark features, it is our basic facial features. It includes the forehead, the eyes and the part in between our eyebrows.
I hope this clears your doubt, thank you
Last edited 2 years ago by JAIN, Prabhansh
SINGH, Dilsher
January 19, 2022 2:24 pm
Hello Everyone,
Welcome to FeverSENSE. Feel free to ask us questions that intrigue you.
JAIN, Prabhansh
January 19, 2022 1:49 pm
Hello Everyone,
I am Prabhansh Jain a team member of FeverSENSE and I will be here to assist you with all your questions. Feel free to ask anything.
Can your system process video in real-time as of today? If yes, what is the frame rate?
Currently, we have only processed facial recognition for the unmasked datasets and the frame rate is around 17-22fps. However, it is also dependent on the resolution of the webcam you are using.
It will be nice to have the framerates under different situations in your final presentation.
Can you elaborate on the data collection process? Do you think 10,000 images are sufficient?
10,000 images are decent enough to get a minimum level of accuracy however since we are performing facial recognition on two factors unmasked and masked with different points for landmark detection. Nearly 30,000 images in batches would be more ideal. Each batch is divided into 3 sets validation, training, and testing in which validation and training set keep changing with each forward and backward propagation so we can train it through more random collection. Thus cross-validation on batches of data containing 30,000 images was ideal for us.
Awesome job guys!
Thanks Anushka
In the poster, you mentioned splitting the data set into a training set and a validation set. Do you mean a training set and a testing set? There seem to be a lot of hyper-parameters to tune. Why is the split not the usual combination of a validate set, a training set, and a testing set?
Hi Yip,
Yes I agree by what you are saying and yes the split has been done into three sets. A validation set, a training set and a testing set which is the common split.
Thank you for spotting the error. We will change it in the poster.
Is there a pair of color and thermal cameras for each algorithm? Are the algorithms running simultaneously and being selected/aggregated to form the output?
Hi Yip,
Thanks for your question, yes there are two pairs of coloured and thermal cameras to produce one single output, and yes the algorithm will be an infinite functional loop. This means on every new person detected the algorithm will keep running itself to get the best results.
Hope that answers your question, thank you.
I do not understand the part that “algorithm to select camera depends on accuracy” and “The existence of
multiple cameras
allows multiple outputs
which will be compared on another algorithm and the most
accurate result will be displayed on a monitor”. Would you mind explaining them further? Will there be multiple algorithms running simultaneously in the final system?
Yes, there will be multiple algorithms running however but will be sequential so face recognition is the first followed by the mapping of the pair of thermal cameras and RGB colored cameras, and finally the selection algorithm. Here we look at the most accurate results on the facial landmark, however this is still in development phase so currently we don’t have the data to support this through empirical evidence. The idea is to select the most accurate temperature through facial landmark temperature mapping, our idea basically checks the set of facial landmark checkpoints and sees the one with the minimum difference from each other so we can select that pair of colored and thermal cameras.
It will be nice to have a clearer flow chart in your final presentation.
We will work on that but due to covid restrictions, our entire team is not available to work on the hardware so we are doing multiple objectives not in proper order so currently uncorrelated tasks are being done not in sequence, therefore, I can understand your confusion. We will for sure have more coherent progress when we integrate all those sections. Thanks for your comments
Hello Mr. Singh,
Could you tell us some technical challenges your team faced in this project and how you overcame them. Thanks!
Regards,
Param Khurana
Hello Mr. Khurrana,
We have divided our work into multiple objectives. In hardware installation, the first is the driver required to be able to read data from the JTAG USB else an error regarding detection would be there. These are drivers for the different OS on ftdichip, this will allow the JTAG pod to be visible. Secondly, during the landmark facial detection the lack of our program to use masked individuals presented an obstacle for which we had to increase the dataset for our pre-trained machine learning model. Currently, we are facing issues regarding the mapping between the pixels of the coloured camera and thermal camera so we can accurately read the temperature of the facial landmarks featured detected during the facial recognition step.
References
FTDICHIP – http://www.ftdichip.com/Support/Documents/InstallGuides.htm
Masked Dataset – https://www.kaggle.com/getting-started/182585
Regards,
Dipankar Sharma
Good afternoon, I am student from group WA01.
I saw that there is an objective statment ” To identify the Optimum angle placements of the Camera to maximize the range and enhance the detected results”
Will the algorithm calculate the optimum angle placement? Or this will be calculated/found by the person who setup the system?
Also, may I have more detail on the “landmark features” such as images of them?
I guess I saw similar images on other group that also supervised by Professor Albert Wong. But it seems it was not shown on your poster.
Thanks a lot
Thanks for your question. The configuration will be manual however we will check the accuracy through the detection of landmark features from the AI algorithm and the setting with the most accurate temperature will be selected.
We did not mention the landmark features in our poster as we did not reach this part when we made the poster. However, now we have completed it and if you would like to see it, please send me your email details if possible as I cannot attach images on this comment.
Thanks!
Thanks for both of your reply, I guess you can use services such as imgur to post the link of your image files or even new version of poster.
So that other viewers can also take a look on your updated results.
Yes we will surely keep that in mind, thanks for the suggestion
Hi Kai Hang,
Thanks for your question. For the angle, it will be set manually but there will be an algorithm that would check if all the landmark points are being identified and can detect fever. Coming to the landmark features, it is our basic facial features. It includes the forehead, the eyes and the part in between our eyebrows.
I hope this clears your doubt, thank you
Hello Everyone,
Welcome to FeverSENSE. Feel free to ask us questions that intrigue you.
Hello Everyone,
I am Prabhansh Jain a team member of FeverSENSE and I will be here to assist you with all your questions. Feel free to ask anything.
Thank you