How ushasi/siamese-spatial-graph-convolution-network works
Project type and purpose
• This is a research‐oriented Python/Matlab‐data pipeline implementing a Siamese graph convolutional network (GCN) plus an MLP for remote sensing image retrieval.
• The README confirms:
• It loads precomputed graph features from PatternNet/UC-Merced datasets
• It trains a Siamese MLP on top of those graph features
• It reports retrieval MAP and precision metrics
File structure
• Top level:
• .mat files (X_train.mat, y_train.mat, pattGCN.mat, etc.) act as data sources.
• main.py orchestrates end-to-end pipeline.
• input_data.py handles reading .mat files and splitting train/test.
• GCN-singlelabel/graphcnn:
• layers.py, network.py, network_description.py define the core GCN architecture.
• helper.py contains utility functions.
• experiment_multilabel.py runs experiments on multilabel data (evaluation and logging).
• GCN-singlelabel/graphcnn/setup:
• helper.py, ucmerced.py and ucmerced_new.py build adjacency matrices and feature graphs from raw datasets.
README for additional insights
• Data → Graph Construction → GCN feature extraction → Siamese MLP → Retrieval/evaluation.
• Key stack: Python, TensorFlow, Matlab data.
Detail
• Every script group: show graphcnn/setup as a submodule under Preprocessing
• That experiment_multilabel.py consumes both the trained model and test split