Type1b Multi-View Spatial Reasoning

Type1b contains gravity-settled Blender scenes, 26-view RGB renders, auditable spatial QA, and raw model predictions. Four vertical-spread conditions are released under the uniform names zspread_025, zspread_050, zspread_075, and zspread_100.

Generation, evaluation, and analysis code is available in the DM3SR GitHub repository.

Dataset layout

scenes/zspread_XXX/                 scene JSON and validation metadata
physics/zspread_XXX/                settled physics reports
renders/zspread_XXX/                26-view PNGs and render_report.json
qa/zspread_XXX/                     canonical QA JSONL
qa_ablation/view_count/zspread_XXX/ paired 2--6 input-view controls
evaluation/zspread_XXX/             model predictions and scores
evaluation/view_count/zspread_100/  input-view-count predictions
metadata/evaluation_lineage.json    QA/input-image lineage audit

The main tasks are object_counting, relative_distance, and relative_position. QA rows reference images relative to their corresponding renders/zspread_XXX directory.

Canonical render counts

Every released scene directory also contains render_report.json. The scene specification collection contains 50 JSON files per level; Blender physics and render validation determine the released render/QA subset.

External assets

Scene specifications identify OmniObject3D assets by category and object ID. Raw OmniObject3D scans and Poly Haven HDRIs are not redistributed in this dataset; obtain them under their original licenses. The companion GitHub repository contains category filters, HDRI checksums, generation, rendering, QA, evaluation, and analysis code.

Release integrity

Use type1b/manifests/huggingface_dataset_manifest.json in the companion code repository for the exact local-source to Hub-destination mapping. The canonical selection follows the QA report_path lineage rather than historical folder labels, which avoids uploading an earlier unused z-spread render generation.

Some retained historical model runs evaluated the same counting questions with an earlier random input-view selection, and some Qwen3 z=1.0 spatial files use historical question IDs. Predictions embed their actual question and image list. metadata/evaluation_lineage.json labels every file as exact, ID-remapped, or same-content/different-input-views so these conditions are never conflated.