-
Clone the repository and in the project directory run:
cd AutoTend -
Create and activate a virtual environment (A working installation of python3.11 is required on target machine):
python3.11 -m venv venv source venv/bin/activate -
Install dependencies:
pip install -r requirements.txt
-
Run the server using your machine’s IPv4 address (not
127.0.0.1orlocalhost):python manage.py runserver <your-ipv4-address>:<port>
⚠️ Note: To get your IPv4 address in Python:
import socket
s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
s.connect(("8.8.8.8", 80))
print(s.getsockname()[0])
s.close()We conducted two experiments to evaluate the performance of different face detection models and to assess whether intra-lecture clustering improves identification accuracy.
Methodology:
We compared three face detection models: Buffalo L, Buffalo Sc, and Antelope v2. Each model was evaluated by computing the average F1 score across all pre-midsem lecture recordings. The experiment was conducted over various combinations of face detection and face similarity thresholds.
- Face detection threshold refers to the confidence score for whether a detected bounding box actually contains a face.
- Similarity threshold refers to the cosine similarity score between two face embeddings; it determines whether two faces are considered a match.
Pipeline Architecture:
For each lecture, a model was run on the faces detected from all photos. For every detected face, we computed the cosine similarity against all ground truth embeddings. If the highest similarity score for any ground truth face exceeded the specified threshold, that student was marked as present.
Below are the F1 score heatmaps for all three models:
Result:
Antelope v2 achieved the highest average F1 score across the evaluated thresholds, with a peak F1 score of 93.14% at a 0.25 similarity threshold and a 0.3 face detection threshold, making it the most effective model among the three.
During each lecture, the TA captures 7–8 photos. We tested whether clustering detected faces within a lecture and applying majority voting (based on ground truth matching) would improve accuracy. This experiment used Antelope v2, based on its performance in Experiment 1.
Result:
Using intra-lecture clustering, the F1 score dropped from a maximum of 93% to 86.04%. Hence, we discarded this technique and opted to compare each face individually to the ground truth across all faces.
Upon accessing the server at http://<your-ipv4>:<port>, you will be greeted with the login screen.
The default login credentials are:
- Username:
0 - Password:
password
This can be changed manually in the database.
⚠️ Note: For the current submission, the ground truth has already been added to the database. You can always click onresetbutton on the homepage and upload from your end if you want to test.
To upload the ground truth, follow these steps:
- Prepare a zip file containing one image (not I-Card image) of each student in .jpg. The filename should be the student's roll number (e.g.,
12345.jpg).
- The images should not be in any subdirectories.
- On the home page, if the ground truth has not yet been uploaded, you will be prompted to upload the zip file.
- Select the zip file and click Save Changes.
⚠️ Note: During this time, to avoid data consistency issues, you won't be able to add a class or reset the program. The processing of the ground truth will take approximately 5 minutes. Once completed, the respective buttons will be re-enabled after you refresh the page.
To run inference and add a new lecture, follow these steps:
-
On the home page, click Add New Lecture.
-
Click Select Photos and choose the photos to upload.
-
After selecting the photos, click Upload Photos.
⏱️ Note: Processing 8 images at once typically takes less than a minute.
⚠️ Note: Images should be in .jpg format.
To use the system on your mobile device:
-
Open the provided IPv4 address in Safari.
-
Tap the Share button and then select Add to Home Screen.
Now you can access the system like a native app on your mobile device.
