How whiffe/scb-dataset works
Project type and purpose
• This is a research-oriented, modular Python toolkit for detecting and analyzing student classroom behaviors.
• It contains data-preprocessing scripts, model-training code (YOLOv7+Biformer, LSTM), evaluation modules (rule-based & neural), similarity-calculation utilities, and visualization tools.
File structure
Top-level directories define major functional areas:
• Data Preparation
– yolo2ava/* (video→frame extraction, AVA annotation conversion)
– SCB_extend_ava/img2video.py (extend dataset via videos)
– FPS_IOU_png2jpg_yaml/* (image format conversion, YAML configs, FPS & IOU benchmarks for YOLO variants)
• Model Training & Inference
– YOLOv7+Biformer/* (detection code, model scripts)
– EvaluationMethod/LSTM/* (Jupyter notebooks & Python scripts for sequence modeling)
• Rule-based & Statistical Evaluation
– EvaluationMethod/Rule-based assessment…/* (behavior rule scripts, CSV data merges, down-video generation)
• Similarity Calculation
– Behavior-Similarity-Calculation/*.py (compute behavioral‐sequence similarity, image filtering)
• Analysis & Visualization
– Student_Behavioral_Concentration_Analysis/* (Excel‐driven analysis scripts)
– figures/*.py (stacked bar chart generator)
– yolo-gradcam/*.py (heatmap visualization for YOLOv5/7/8)
Extract additional insights from the README
• Focus is on the SCB-dataset and related open datasets (STBD-08, ClaBehavior, etc.)
• Emphasis on dataset links, paper citations, dataset statistics
• No web service or front-end—pure research pipelines
1. External Dataset Sources (Baidu, Roboflow, paid websites)
2. Data Ingestion & Conversion
• yolo2ava, SCB_extend_ava, FPS_IOU_png2jpg_yaml (png2jpg, YAML gen)
3. Data Storage
• Video files, frame directories, CSV/Excel annotation stores
4. Model Training Modules
• YOLOv7+Biformer, LSTM notebooks
5. Inference & Detection Output
• Trained weights, detection logs, JSON/CSV outputs
6. Evaluation & Metrics
• Rule-based folder, LSTM eval, behavior-similarity calculation
7. Analysis & Visualization
• yolo-gradcam heatmaps, figures/*.py charts
8. Results & Reports
• Summary CSVs, generated down-videos, Excel sheets
– External Datasets → Data Ingestion
– Ingestion → Data Storage → Model Training
– Trained Models → Inference → Detection Output
– Detection Output → Evaluation Modules → Metrics
– Metrics → Similarity Calculation & Analysis → Visualization
– Visualization → Final Reports
c. Architectural patterns & principles:
• Pipeline pattern (linear data flow from raw inputs to reports)
• Modular separation: preprocessing, modeling, evaluation, visualization
d. Key technologies & frameworks:
• Python scripts & Jupyter notebooks
• YOLOv7 (PyTorch) + custom Biformer head
• LSTM implementations (Keras/PyTorch)
• Rule-based Python logic for behavior assessment
• OpenCV / ffmpeg (video↔frame)
• Pandas / Excel for analysis
• Matplotlib / Grad-CAM for visualization
Detail
• Every folder/script group as its own component.
• Intermediate data artifacts (frames, YAML configs, CSVs).
• External dataset sources and model-weights repositories as boundary elements.