Welcome to CompVis, our project in implementing a Deep Learning model to detect and recognize faces from video files applied to industrial safety.

	Kolapally Sai kalyan	GitHub	Portfolio	LinkedIn
	Daniel Osório	GitHub	Portfolio	LinkedIn
	Merle Buchmann	GitHub	Portfolio	LinkedIn
	Kranthi Maddishetty	GitHub	Portfolio	LinkedIn

The primary objective of our project was to leverage computer vision and advanced face detection and recognition technology for emergency monitoring in industrial safety contexts. To achieve this goal, we used an episode of the popular TV show "The Office" that depicted a fire drill scenario. We then trained our face recognition model using a dataset of six characters from the show. Our ultimate aim is to create a robust and reliable tool that can help improve emergency response by quickly identifying and tracking individuals.

• Python backend:
  • Image and video manipulation using OpenCV
  • Face detection using MTCNN
  • Face recognition deep learning model using TensorFlow transfer learning with EfficientNetV2 trained with a dataset of 1500 images of each character after image augmentation
• API with FastAPI
• Streamlit frontend

🧪 You can test our app here https://compvis.streamlit.app/ 🧪

Detection and identification of faces on a single image -Select the image option on the left navigation bar, upload an image and the app returns a final image with a face bounding box and predicted face identification -By default the app uses a threshold of 70% probability to consider the identification positive. Bellow that value faces are labeled as unknown	(Click to enlarge)
Detection and identification of faces on a video -Select the video option on the left navigation bar, upload a video, and the app returns a final video with a face bounding box and predicted face identification -By default the app will label the first 30 seconds of the video and will sample faces every second. The app returns a video with a face bounding box and predicted face identification. The labeled frames are duplicated a defined number of times to better visualize them	(Click to enlarge)
See a quick demo of this app here

This project was developed as part of the Le Wagon Data Science Bootcamp Batch 1181 Online (Feb-Mar2023) over two weeks