Emotion Detection Model Comparison

Job ID: 40099748

Budget: $10 – $30 USD

I have a collection of pre-recorded classroom videos featuring adult students (20+). The task is to build two separate pipelines that estimate emotion from both facial cues and body language, then benchmark their performance:

• Pipeline 1 – CNN-BiLSTM: extract spatial features frame-by-frame, pass temporal sequences through a BiLSTM, and classify the five basic emotions (happiness, sadness, anger, fear, surprise).

• Pipeline 2 – YOLO-based: detect faces and upper-body regions with YOLO, feed cropped sequences to an emotion head, and output the same five-class predictions.

After training on the videos, compare accuracy, F1, and real-time throughput. Once the stronger approach is confirmed, deploy it for live testing (webcam or external camera) so emotions can be displayed or logged in real time.

Deliverables must include:
– Clean, reproducible code (Python; TensorFlow or PyTorch; OpenCV where needed)
– A concise report detailing architecture choices, hyper-parameters, evaluation metrics, and the head-to-head comparison
– A lightweight deployment script or notebook that runs on a standard laptop and shows live predictions at usable FPS

Acceptance is based on: minimum 80 % overall accuracy on a held-out set, real-time inference ≥10 FPS, and clear documentation that allows me to retrain or fine-tune the models later.