all projects
pytorchcomputer-visiontensorrt
Facial Expression Recognition
ResNet + U-Net hybrid for real-time emotion detection, optimised with TensorRT for video. Highest accuracy on FER2013 in the cohort.
What it is
A PyTorch model combining ResNet + U-Net masking blocks for facial-expression classification on FER2013, then optimised for real-time video inference.
What I shipped
ResNet + U-Net Hybrid
Designed and trained custom neural net with segmentation masking overlays, reaching cohort-highest 77% accuracy on FER2013.
TensorRT Optimization
Compiled model graph for CUDA execution to optimize inference path latency for real-time video classification.
LFFD & OpenCV Pipeline
Fused OpenCV face tracking with LFFD detector to crop and preprocess target faces rapidly in the input frame stream.
Key Metrics
- FER2013 accuracy
- 77%
- venue
- BCS · IIT Kanpur
Technology Stack
pytorchcomputer-visiontensorrt