| --- |
| pretty_name: Youtube Self Depth QA |
| task_categories: |
| - visual-question-answering |
| tags: |
| - spatial-reasoning |
| - depth-estimation |
| - stereo-vision |
| - image-to-text |
| license: other |
| configs: |
| - config_name: level1 |
| data_files: |
| - split: train |
| path: data/level1/train.parquet |
| - split: test |
| path: data/level1/test.parquet |
| - config_name: level2 |
| data_files: |
| - split: train |
| path: data/level2/train.parquet |
| - split: test |
| path: data/level2/test.parquet |
| - config_name: level3 |
| data_files: |
| - split: train |
| path: data/level3/train.parquet |
| - split: test |
| path: data/level3/test.parquet |
| --- |
| |
| # Youtube Self Depth QA |
|
|
| > **Release status: private review only. Do not make this repository public until every source video/image has a documented redistribution right and the Level-3 validator is independently completed.** |
|
|
| ## Dataset summary |
|
|
| Youtube Self Depth QA is a visual spatial-reasoning benchmark generated from relative disparity maps. It has three task levels: |
|
|
| - **Level 1:** identify nearest/farthest marked point, optionally relative to a reference point. |
| - **Level 2:** select the correct ordering of four marked points. |
| - **Level 3:** choose the trajectory whose depth changes monotonically in the requested direction. |
|
|
| Each QA sample is rendered at six intervention baselines (`0`, `0.05`, `0.10`, `0.15`, `0.2`, `0.25`). This release represents each `(sample_id, baseline)` pair as one row. The `image` column is embedded in Parquet for Dataset Viewer compatibility. |
|
|
| ## Performance across stereo baselines |
|
|
|  |
|
|
| The figure reports accuracy under the six stereo-baseline settings for Qwen2.5-VL-3B, Qwen2.5-VL-7B, Qwen3.5-4B, and Qwen3.5-4B-Base across VSR and the three dataset levels. Dashed horizontal lines mark chance performance for the corresponding task. |
|
|
| ## Splits |
|
|
| The train/test split is grouped by source video clip, rather than by individual frame. This avoids placing near-duplicate frames from a clip in both splits. The original split manifest reports 34 clips in total, 6 held-out test clips, and about a 10.8% test fraction for each level. |
|
|
| ## Fields |
|
|
| `image`, `sample_id`, `baseline`, `level`, `task_type`, `question`, `answer`, and `metadata`. `metadata` is a JSON string containing the geometric information used during generation; host-specific source paths are intentionally omitted. |
|
|
| ## Important limitations |
|
|
| - Depth/disparity values are relative and are not metric 3D depth. |
| - Labels are generated from estimated disparity and can inherit its failures, especially around occlusions, reflective surfaces, thin structures, and moving content. |
| - This dataset contains imagery derived from online videos. The publisher must document the source, creator, applicable license, access date, and redistribution right for every clip before releasing it publicly. |
| - Images may contain identifiable people, trademarks, or other sensitive visual content. Review, removal procedures, and applicable privacy requirements must be documented before public release. |
| - The current repository's generic verifier covers Levels 1–2 only. Level 3 requires a separate independent validator before a public benchmark claim. |
|
|
| ## Intended use |
|
|
| Research on visual spatial reasoning, disparity-aware QA, and evaluation of baseline-conditioned image understanding. It is not intended for safety-critical depth estimation, identity-related inference, or use as a source of metric depth ground truth. |
|
|
