JaweriaAsif745/Human-Activity-Recognition
Real-time Human Action Recognition using CNN + LSTM in PyTorch. Recognizes actions from videos or webcam, displaying top predictions with confidence scores via a Streamlit web app.
What's novel
Real-time Human Action Recognition using CNN + LSTM in PyTorch. Recognizes actions from videos or webcam, displaying top predictions with confidence scores via a Streamlit web app.
Code Analysis
6 files read · 3 roundsA Streamlit web app that runs inference on a CNN+LSTM action recognition model (ResNet18 + LSTM) over video uploads or webcam frames, displaying top-k predictions with confidence scores.
Strengths
The model architecture is clean and well-structured, and the inference pipeline handles both video file and webcam inputs with reasonable frame sampling and transform logic. The code is readable and follows standard PyTorch patterns.
Weaknesses
No training code is included, the model checkpoint is a 133-byte LFS placeholder (not real weights), there are zero tests, and the webcam loop is a blocking while-loop inside Streamlit that will freeze the UI. The README claims features like 'checkpointing' and 'save/resume training' that don't exist in the code.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
38
Contributors
1
Files
9
Active weeks
13
Repository
Language
Python
Stars
1
Forks
0
License
MIT