Inference and evaluation code for When Size Helps and Hurts: Physical Measurements in 3D Question Answering.
The pipeline assembles cached object, image, and directed-relation features, runs a frozen 3DGraphLLM under measurement controls, scores answers, and computes paired statistical comparisons.
Supply prepared features, physical measurements, question annotations, and pretrained model checkpoints. ULIP-2 features also require a fitted bridge checkpoint. Feature extraction and bridge fitting precede this pipeline.
From the repository root in PowerShell or a Linux shell:
python -m pip install -r requirements.txt
python -m unittest discover -s experiments/scale_interface/testsSee the run commands for preparation, inference, scoring, and analysis, and the experimental protocol for input formats and controls. The implementation and editable configuration examples are in experiments/scale_interface.