Deep learning based visual trackers

[Muhammad Alim]

Hey, hope you are well.
I am Muhammad Talha Umair Alim.
I am currently a junior year student in Bachelor's of Computer Science at National University of Sciences and Technology (NUST), Islamabad. I am passionate towards image processing and deep learning. I have been learning deep learning since my last year's summer vacations. I am not expert or have years of experience in computer vision but I am confident and hard working enough to get this project nicely done.
SiamMask is one of the best model architecture that can be used for object detection.
As per skills requirement, I have studied courses Linear Algebra, Calculus I, Calculus II at my university. Moreover, I like programming and have good experience in it, you can check my github profile: https://github.com/talha3111997

[Ilya Elizarov]

Hello!
My name is Ilya, I am one of the mentors of "DL based visual trackers".
I am glad to see that you interested in our project, but I have several questions:
1) Could you tell more about your skills? (math, programming)
2) Could you tell more about your experience in DL?
3) Do you know any other state-of-the-art trackers?
Note: our project require knowledge of Python and C++.
If you have any questions - just ask, I will answer as soon as possible.

[Muhammad Alim]

1) Could you tell more about your skills? (math, programming)

        As far as programming is concerned, you will be glad to know that I love programming and if you want to know my logic and experience, kindly have a look at my github profile repositories: https://github.com/talha3111997 

and secondly I have knowledge of Calculus, Calculus II and Linear Algebra as I have studied them in my university (I don't know how can I tell and assure you about my math knowledge. I can share my transcript with you so, you can view my grades in these math courses or if you want me to share specific topics about math upon which I have strong grip. Just let me know )

2) Could you tell more about your experience in DL?

     You can check my Neural Networks and Deep learning course certificate here: https://www.coursera.org/account/accomplishments/certificate/N8CH8S9ETHM4?utm_medium=certificate&utm_source=link&utm_campaign=copybutton_certificate.

     

        I have not only a experience in transfer learning in vision rather I have also implemented fully connected neural networks from scratch in octave as well as in numpy to get the solid concept and still I have been learning to become better and better.

    You can look some of my work in Deep Learning here: https://github.com/talha3111997/Pytorch-Projects

3) Do you know any other state-of-the-art trackers?

      CenterNet, YOLO, SSD, Cascade R-CNN are some of the state of the art trackers but the best choice depends on the preference that either we prefer inference speed or detection accuracy.

 

Note: our project require knowledge of Python and C++.

   Yes, I have worked and used both languages (Python and C++) for different purposes. I have also uploaded some of my work(Data Structures and Algorithms) which I did in c++ on Github. You may check here: https://github.com/talha3111997/Data-Structures-and-Algorithms

  For python you can check my following repository:

   https://github.com/talha3111997/Image_Processing

[Muhammad Alim]

Dear Mentor, can you please tell that upon which basis we should choose state of the art tracker i.e inference speed or detection accuracy.
Thank you

[Muhammad Alim]

Respected Mentor, one more thing I want to make clear is that we need to choose the model architecture based on its performance in Real time?

[Muhammad Alim]

Respected Mentor, opencv is already providing some best trackers like SSD, Faster-RCNN, YOLOv2, tiny YOLO, YOLOv3 through different frameworks like tensorflow. So, do we need to implement state of the art tracker other than these?

Just like some models are mention here: https://github.com/opencv/opencv/blob/master/samples/dnn/models.yml

[siddhant kapil]

The models that you have mentioned are object detection models and they are different from tracking models. These models need data to train on specific classes and during inference, they give the class and location of the object. While tracking models like ATOM, and DiMP they will take the location of the object in a frame and then track the object in subsequent frames. These models just give the location of the objects in upcoming frame but don't classify them.

[Muhammad Alim]

oh alright. Thank you so much.

And which model OpenCV is currently using?

[siddhant kapil]

Medianflow, KCF, TLD, MIL, Boosting