vsg_eval

Evaluation harness, annotations and result dumps for video scene-graph work — the measurement side of the SVG2 pipeline. This is a working research tree published as-is, not a packaged library.

Contents

Path	What
evaluate.py, metrics/, utils/	the evaluation entry point and metric implementations
annotations/	evaluation annotations, including the PVSG set
dataset_quality/	dataset-quality study: demo renders and cached intermediates (the bulk of the files)
toksel_eval/	token-selection evaluation, including LLM-judge scripts (sync and Batch API)
pipeline_eval/, sam2Pipe_eval/	end-to-end pipeline evaluation runs
eval_pvsg/, eval_1020/, eval_internvl3_1b/, eval_qwen2.5_baseline/	result dumps per model/config
OpenPVSG/	OpenPVSG-derived evaluation code
release_example/	worked example

Notes for anyone using this

• Paths are absolute cluster paths. Scripts reference a filesystem layout that will not exist for you; expect to rewrite roots before anything runs.
• API keys were removed. The LLM-judge scripts under toksel_eval/pvsg/ and annotations/pvsg/gpt_score.py previously carried hardcoded OpenAI keys. Each now reads os.environ.get("OPENAI_API_KEY", "") — export your own key to run them.
• dataset_quality/ is mostly cache. ~48k files of demo renders and cached intermediates from a quality study. Useful for inspecting what was measured; not required to run an evaluation.
• No model weights here. See royguw/ft_qwen3_5 for the baseline training code and royguw/grounding-data / royguw/mm-olmo-images for data.

Licensing

Evaluation code and result dumps here are released for research use. Annotations and code derived from upstream projects — OpenPVSG and the PVSG benchmark in particular — remain under their original licenses, and evaluated model outputs are subject to the terms of the models that produced them. Check the upstream license for any component you reuse.