RECONSTRUCTION.md

Reconstruct the GQA evaluation inputs locally

The public candidate includes our labels, factual source IDs, image/question checksums and our edit deltas for all 3,000 GQA views. It omits upstream questions, full semantic programs, full graphs and photographs because an explicit GQA annotation redistribution grant was not found. Read the upstream terms when obtaining the source files. Do not upload the reconstructed directory as part of this candidate without resolving the missing grant.

1. Obtain questions1.2.zip and sceneGraphs.zip from the official GQA downloads. Extract val_balanced_questions.json and val_sceneGraphs.json.
2. Obtain the official images.zip, or the equivalent original images from Visual Genome. Put the selected original JPEGs in one directory named {image_id}.jpg. The needed 1,000 IDs are in the gqa configuration and reconstruction recipes. Re-encoded third-party mirror images may not reproduce the frozen bytes.
3. Install the pinned requirements.txt and run:

python scripts/reconstruct_hf_gqa.py \
  --questions /path/to/val_balanced_questions.json \
  --scene-graphs /path/to/val_sceneGraphs.json \
  --images /path/to/original-images \
  --output ../gqa-reconstructed

The script performs no network download or upload. It requires a new output directory outside the release. It checks original question strings and source programs, canonical graphs, normalized photograph bytes, and every final image against the frozen evaluated identities. It replays the source program and the symbolic alternatives, then writes an ImageFolder-compatible metadata.jsonl, 3,000 JPEGs and the reconstructed program/edit evidence. Any mismatch fails explicitly; it does not substitute newly generated examples.

from datasets import load_dataset
gqa = load_dataset(
    "imagefolder",
    data_files={"test": [
        "../gqa-reconstructed/images/*.jpg",
        "../gqa-reconstructed/metadata.jsonl",
    ]},
    split="test",
)

Use this explicit file/split selection. Directory-wide split inference can pick up validation-report filenames; those reports are not model inputs. These 3,000 examples belong to the benchmark's evaluation-only test cohort.

The GQA renderer decodes the source JPEG to RGB, writes a normalized JPEG at quality 88 with 4:2:0 subsampling, and applies a (127,127,127) patch over one recorded object box for edited views. FULL avoids a second JPEG encoding. Patched views are encoded once from the normalized source. Exact codec behavior is why the pinned Pillow version matters. CLEVR and PlotQA need no reconstruction for model evaluation because their evaluated pixels are embedded in Parquet.

The CLEVR source worlds/programs and renderer settings are bundled for auditing. Optional regeneration of the rendered CLEVR scenes additionally requires the official generator/assets at revision f0ce2c81750bfae09b5bf94d009f42e055f2cb3a, Blender 2.79b/Cycles and the settings in the view evidence. These dependencies are not needed to load the dataset or to replay its source-program labels. GQA graph alternatives are symbolic scene evidence and do not establish identical pixels for hypothetical photographs.