How fvl2020/2dimage2bmi works
Project type and purpose
• This is a full-stack research codebase for estimating human weight (BMI) from a single 2D image.
File structure
Top-level:
• ALL_feature/ – CSVs of precomputed features (train & test).
• bodyfeature/ – placeholders for anthropometric computations.
• modeling/ – PyTorch model definitions (resnet, segmentation module, affine alignment, skeleton feature extractor, build_model.py, templates.json).
• model.py – likely an entry point to assemble and run the network.
• util.py – shared utilities.
Key scripts:
• Data pipeline: Data.py (I/O), Datasets_BFDF.py (custom Dataset).
• Preprocessing/Extraction: Detected.py (object detection via Detectron2), Human_Parse.py (human parsing), BodyFeatureExtractor.py, DeepFeatureExtractor.py.
• Aggregation & regression: CombineAndRegress.py, OtherTypeNet.py.
• Lifecycle: Training.py (training loop, tensorboardX logging), Regression.py (inference/test), ANOVATest.py (statistical analysis).
README for additional insights
• Describes dual-branch architecture.
• Lists dependencies: PyTorch, CUDA, Detectron2, scikit-learn, scipy, pandas, OpenCV, tensorboardX.
• Points to pretrained models (downloaded from Google Drive).
• Shows result plots, cites the IEEE TMM paper.
Final recommendations
• Be very thorough: capture all Python scripts, modules, and data assets as distinct nodes.
• Don’t overcomplicate—focus on separating the project into logical components matching the file structure and README description.
• Reference the existing framework/framework.jpg as a starting template; expand it to show data I/O, training loop, and evaluation stages in detail.