VQA README: shared page layout

vqa/README.md

@@ -1,226 +1,127 @@
 # VQA ground truth
 
-Machine-generated VQA ground truth over RPX scene data, covering all 100 scenes at full frame density (every frame of every mos phase and the ego sequence). Three configs, one per task:
-
-| Config | Rows | Task |
-|---|---|---|
-| `attribute.parquet` | 3,241,564 | "which object is X" -- color/material/function/composition/count/odd-one-out/synonym/absence, from each object's real catalog attributes |
-| `spatial_binary.parquet` | 744,741 | left-right / above-below yes-no between two named objects |
-| `spatial_bbox.parquet` | 543,017 | "which object is X" answered with a bounding box -- closest to camera, farthest from a reference, leftmost/rightmost |
-
-Spatial configs are `mos`-only (ego has no depth map). Generation code, task definitions, and correctness guarantees (uniqueness checks, depth-unit handling, phase-wide dedup for attribute facts) live at `data/mask_annotation/visual_grounding_gt/vqa_gt/` on the `itay/vqa` branch of the code repo.
+Machine-generated question-answer pairs over all 100 RPX scenes, at full frame density: every frame of
+every Clutter, Interaction and Clean phase and of the Ego sequence. The single-image files back task T5
+(VQA) and the two-image files back task T6 (In-context VQA). Answers come from each object's real catalog
+attributes and from metric depth.
+
+To score a model, use the [benchmark subset](metadata/benchmark_subset_v1/BENCHMARK_SUBSET_README.md)
+rather than these full banks.
+
+<p align="center">
+  <a href="#files">Files</a> ·
+  <a href="#single-image-questions">Single-image questions</a> ·
+  <a href="#in-context-questions">In-context questions</a> ·
+  <a href="#schema">Schema</a> ·
+  <a href="#validation-and-provenance">Validation and provenance</a>
+</p>
+
+## Files
+
+| File | Rows | Questions |
+|:--|--:|:--|
+| `attribute.parquet` | 3,241,564 | which object has a given attribute |
+| `spatial_binary.parquet` | 744,741 | left/right and above/below, yes or no |
+| `spatial_bbox.parquet` | 543,017 | spatial questions answered with a box |
+| `incontext_mos_attribute_bbox.parquet` | 317,873 | attribute transfer, MOS |
+| `incontext_ego_attribute_bbox.parquet` | 113,193 | attribute transfer, Ego |
+| `incontext_mos_spatial_bbox.parquet` | 318,017 | farthest from a reference, MOS |
+
+`metadata/incontext_v1/` holds the in-context release record: `VALIDATION_REPORT.json`,
+`FUNCTION_ALIAS_DECISION.md`, `SHA256SUMS` and `RELEASE_MANIFEST.json`. The in-context Parquets are
+full-density question banks, with every eligible fact turned into a row; the sampled benchmark lives in
+`metadata/benchmark_subset_v1/`.
+
+## Single-image questions
+
+Attribute questions ask which object matches a color, material, function or composition, which one is
+the odd one out, how many objects share an attribute, which object a synonym refers to, or whether an
+object is absent. Spatial questions compare two named objects (left/right, above/below) or ask for the
+object closest to the camera, farthest from a reference, or furthest left or right.
+
+Spatial files cover MOS only, because Ego frames have no depth map. Generation enforces unique answers,
+handles depth units explicitly, and deduplicates attribute facts across each phase. The generator is in
+[`data/mask_annotation/visual_grounding_gt/vqa_gt/`](https://github.com/IRVLUTD/RPX/tree/main/data/mask_annotation/visual_grounding_gt/vqa_gt)
+in the code repository.
+
+## In-context questions
+
+Every in-context row asks about a pair of images:
+
+1. Image 1 (`reference_image`) is a deterministic crop of one reference object from its single-object
+   scan (SOS).
+2. Image 2 (`target_image`) is a MOS or Ego frame from the single-image files, with the same
+   `scene_id`, `kind`, `phase` and `frame` meaning.
+3. `answer_bbox` is `[xmin, ymin, xmax, ymax]` in Image 2's original pixels. `img_w` and `img_h`
+   describe Image 2, never the crop.
+
+| Type | Question |
+|:--|:--|
+| `inctx_attr_single_color` | same color as the object in Image 1 |
+| `inctx_attr_single_material` | same material |
+| `inctx_attr_single_function` | same purpose |
+| `inctx_attr_composition` | same material and purpose |
+| `inctx_attr_odd_one_out` | not made of the same material |
+| `inctx_spatial_farthest` | farthest from the object in Image 1 (MOS only) |
+
+The five attribute types show a different object in Image 1, one that shares the asked attribute with
+the answer. A candidate reference is rejected if any of its other attribute values also belongs to
+another object visible in Image 2, so each question has exactly one defensible answer. References are
+chosen deterministically by a SHA-256 ranking. The spatial type instead shows the anchor object
+itself, which is expected to appear in Image 2.
+
+For matching and ambiguity checks only, `gray` and `grey` count as the same color, compound materials
+such as `wood/metal` are split into their parts, and `to drink`/`drink`, `to eat`/`eat` and
+`to snack`/`snack` count as the same function. Stored text is never rewritten.
+
+Because the published attribute data records some functions under both spellings, 426 rows of
+`incontext_mos_attribute_bbox.parquet` (0.13%; none in Ego) describe the same reference, target and
+frame twice. Both rows are kept, since each pairs exactly with its own published source row.
+[`FUNCTION_ALIAS_DECISION.md`](metadata/incontext_v1/FUNCTION_ALIAS_DECISION.md) lists them.
+
+Ego has no spatial questions of either kind in this release, single-image or in-context.
 
 ## Schema
 
-Common: `scene_id`, `kind` (`mos`/`ego`), `phase` (`0`/`1`/`2`/`null`), `frame`, `img_w`, `img_h`, `type`, `question`, `answer`.
+Every single-image row has `scene_id`, `kind` (`mos` or `ego`), `phase` (`0`, `1`, `2` or null),
+`frame`, `img_w`, `img_h`, `type`, `question` and `answer`. Box questions add `answer_bbox`
+(`[xmin, ymin, xmax, ymax]` in pixels). Depth-based spatial types add `evidence`, a JSON string with
+the measured values behind the answer, and `attr_count` adds `answer_oids` (also a JSON string). Mask
+paths are omitted; `scene_id`, `kind`, `phase` and `frame` identify the frame.
 
-`answer_bbox` (`[xmin, ymin, xmax, ymax]` pixel coords) on bbox-eligible types. `evidence` (JSON-encoded string) on the depth-based spatial types, recording the real values the answer was derived from. `answer_oids` (JSON-encoded string) on `attr_count`. `mask_path` is intentionally omitted (local path, not portable -- `scene_id`+`kind`+`phase`+`frame` already identify the frame).
+In-context rows carry `source_sample_id`, the exact single-image row each one pairs with: same frame,
+type, target object and box. It is the only join key the release stores.
 
----
+<details>
+<summary>Full in-context schema</summary>
 
-## In-context (two-image) VQA -- v1
-
-Three new configs, generated from the single-image configs above rather than
-independently. **These are full-density question banks (every eligible
-published fact turned into a row), not the final ~36K benchmark sample** --
-downstream sampling/capping for a benchmark release is a separate step not
-yet performed.
-
-| Config | Rows | SHA-256 |
-|---|---|---|
-| `incontext_mos_attribute_bbox.parquet` | 317,873 | `86a334247198848390c158537a235937dda2d386c022d56ede3bf135d798a94b` |
-| `incontext_ego_attribute_bbox.parquet` | 113,193 | `398565fc10b58f5beec3f237c5efb8cf0fd18123fed2782f6ab1c15561a9c186` |
-| `incontext_mos_spatial_bbox.parquet` | 318,017 | `6652cba3b471166e4f2dd366d9434495e63a7b80b5271cca61eeceb2ac0ec156` |
-
-Full validation results, the exact-string-ontology decision for function
-aliases, and a per-file manifest (path/bytes/rows/sha256) live under
-`metadata/incontext_v1/`: `VALIDATION_REPORT.json`,
-`FUNCTION_ALIAS_DECISION.md`, `SHA256SUMS`, `RELEASE_MANIFEST.json`.
-
-### The two-image contract
-
-Every row asks a question about a pair of images, not one:
-
-- **Image 1** (`reference_image`) -- a deterministic crop of a single
-  reference object, taken from that object's SOS (single-object-scan)
-  capture. This is what the question means by "the object shown in Image 1".
-- **Image 2** (`target_image`) -- an existing MOS or Ego target frame from
-  the single-image configs above (same `scene_id`/`kind`/`phase`/`frame`
-  semantics as `attribute.parquet`/`spatial_bbox.parquet`).
-- **`answer_bbox`** is always `[xmin, ymin, xmax, ymax]` in **Image 2's own
-  original pixel coordinates** (`img_w`/`img_h` describe Image 2, not
-  Image 1) -- never Image 1's crop coordinates, regardless of question type.
-
-### Question types
-
-Five general (attribute-transfer) types, split across the two attribute
-Parquets by `kind`:
-
-| `type` | Question |
-|---|---|
-| `inctx_attr_single_color` | Which object in Image 2 has the same color as the object shown in Image 1? |
-| `inctx_attr_single_material` | Which object in Image 2 is made of the same material as the object shown in Image 1? |
-| `inctx_attr_single_function` | Which object in Image 2 is used for the same purpose as the object shown in Image 1? |
-| `inctx_attr_composition` | Which object in Image 2 has the same material and purpose as the object shown in Image 1? |
-| `inctx_attr_odd_one_out` | Which object in Image 2 is NOT made of the same material as the object shown in Image 1? |
-
-One spatial type, `mos`-only (inherits the depth-map limitation of the
-single-image spatial configs), in its own Parquet:
-
-| `type` | Question |
-|---|---|
-| `inctx_spatial_farthest` | Which object in Image 2 is farthest from the object shown in Image 1? |
-
-### Reference-selection semantics: Option B vs. Option A
-
-The five general types use **Option B**: Image 1 shows a *different* object
-from the answer, one that happens to share the relevant attribute value(s)
-with the correct answer object in Image 2. A cross-modal, multi-attribute
-ambiguity gate rejects any candidate reference if *any* of its other
-(non-question) attribute values are also independently owned by a different
-object visible in Image 2 -- not just the intended value -- so the question
-has exactly one defensible answer. Reference selection is deterministic
-(sha256-ranked over eligible candidates), not random.
-
-`inctx_spatial_farthest` uses **Option A**: Image 1 shows the *exact same*
-object that the question's "farthest from" anchor refers to -- there is no
-attribute-sharing indirection, and the reference object's presence in
-Image 2 is expected and intentional (it is the geometric anchor the
-distance is measured from), not a violation of any visibility rule.
-
-### `source_fact_id` vs. `source_sample_id`
-
-Every in-context row carries `source_sample_id`, which identifies the exact
-row in the published single-image Parquet (`attribute.parquet` for the
-general types, `spatial_bbox.parquet` for the spatial type) that this row
-was **1:1 paired against** -- same `scene_id`/`kind`/`phase`/`frame`/`type`/
-target object/`answer_bbox` as that published row. This is a strict,
-row-level join key: every in-context row corresponds to exactly one real,
-already-published fact placed in exactly the same frame.
-
-During development, a second, looser identifier -- `source_fact_id`,
-`(scene, kind, phase, type, target_object, normalized attribute value(s))`
-without the frame/bbox scoping -- was used only for auditing (it matches
-whenever a fact was published *somewhere* in the phase, regardless of
-frame, useful for confirming a fact is real before checking whether it also
-landed in this exact frame). It is not a persisted column in these
-Parquets; the frame-exact `source_sample_id` is the only join key the
-release itself carries.
-
-### Canonicalization (matching/ambiguity only, never eligibility or stored fields)
-
-The ambiguity gate treats a small set of attribute-value spellings as
-identical when deciding whether a reference is safe to use, without ever
-altering what is actually stored: `question`, `source_question`,
-`attribute_value`, `attribute_material`, `attribute_function`, and
-`source_sample_id` always preserve the original published text/fact
-verbatim.
-
-- **Color:** `gray` / `grey` are treated as the same color.
-- **Compound materials:** slash- or ampersand-joined compound values (e.g.
-  `"wood/metal"`) are decomposed into their atomic materials before
-  ambiguity is checked, so a reference is rejected if *any* atomic
-  component conflicts with another visible object -- a conservative rule
-  that only rejects more candidates, never fewer.
-- **Function infinitive variants:** `to drink`/`drink`, `to eat`/`eat`,
-  `to snack`/`snack` are treated as the same function.
-
-### Known, quantified, source-inherited semantic duplicates (426 rows)
-
-Because canonicalization is never applied to stored fields, and because the
-published dataset itself independently records some facts twice under two
-literal spellings (e.g. one published row says `"used for drink"`, a
-different published row for the same object says `"used for to drink"`),
-426 rows in `incontext_mos_attribute_bbox.parquet` (0.13% of 317,873; 0 in
-the ego file) end up describing the same Image1-reference/Image2-target/
-frame combination twice, differing only in which spelling of the same
-function was published. Each such row is still a correct, independent 1:1
-pairing to a real distinct published row -- dropping either one would
-break exact source-row pairing, which is treated as the stricter
-guarantee -- so both are kept. Full breakdown, the exact duplicate keys,
-and the "why keep both" reasoning are in
-`metadata/incontext_v1/FUNCTION_ALIAS_DECISION.md`.
-
-### Not yet included
-
-**Ego has no in-context spatial config, and no normal (single-image) ego
-spatial config either** -- both are absent from this release, not just
-this v1 in-context slice. This mirrors the existing single-image
-limitation noted above (spatial configs are `mos`-only because ego frames
-have no depth map); it is called out again here explicitly since it is
-easy to assume in-context spatial covers both `kind`s the way the general
-attribute-transfer types do.
-
-### Provenance
-
-- Generator code: commit `fbf2fd03cd9972d159ac954cfed8fd194deb8d87` on
-  branch `itay/vqa-incontext` of the code repo (same repo/location as the
-  single-image generator referenced above).
-- Every image locator (`target_image`, `reference_image`) is pinned to the
-  immutable Hugging Face dataset revision
-  `93e31d378f1f98eca18a7aa01a2279c9f332440c` (a resolved commit SHA, not
-  the mutable `main` branch name), so assets referenced by these rows
-  cannot change out from under a consumer.
-
-### Validation: exhaustive vs. sampled
-
-Most correctness checks in `VALIDATION_REPORT.json` are **exhaustive**
-(run over every row in the release, not a sample): bbox validity, identity/
-reference-visibility violations, duplicate sample IDs, semantic-duplicate
-detection (both raw-string and function-alias-normalized keys), reference
-asset availability and crop-hash matches, pinned-revision checks,
-source-row pairing against the published Parquets, and per-frame/
-centered-frame coverage across all MOS+Ego cells. The reference-object-not-
-visible-in-Image-2 check is also exhaustive and full-100-scene (not
-sampled), covering every general row; it intentionally does not apply to
-`inctx_spatial_farthest`, whose Image-1 identity is the anchor and is
-*expected* to be visible in Image 2.
-
-The one check that is **sampled, not exhaustive**: an independent
-recomputation of the raw farthest-object distance geometry for
-`inctx_spatial_farthest` against a subset of scenes, used as a spot-check
-in addition to (not instead of) the exhaustive 1:1 row-for-row equality
-proof against the published `spatial_farthest` rows in `spatial_bbox.parquet`
-(every one of the 318,017 rows in `incontext_mos_spatial_bbox.parquet`
-corresponds exactly to one published `spatial_farthest` row -- that
-equality check itself is exhaustive; only the independent from-scratch
-geometry recheck used as a secondary sanity check was sampled).
-
-### Schema (obtained programmatically via `pyarrow.parquet.read_schema`)
-
-All three files share the same top-level column set; the general
-attribute files leave `evidence`/`_dedup_key` null-typed (unused) and
-populate `attribute_kind`/`attribute_value`/`attribute_material`/
-`attribute_function`, while the spatial file does the reverse (populates a
-structured `evidence`, leaves the attribute_* columns null-typed) and adds
-an internal `_dedup_key` column (build-time cap bookkeeping, not meaningful
-to consumers).
+The three in-context files share one column set. Attribute files fill the `attribute_*` columns and
+leave `evidence` and `_dedup_key` null; the spatial file does the reverse. Unlike
+`spatial_bbox.parquet`, the spatial file stores `evidence` as a native Parquet struct, so read it
+directly instead of with `json.loads`.
 
-```
+```text
 scene_id: string
 kind: string
-phase: double                      (0/1/2/null; spatial file uses int64)
+phase: double                  (0/1/2/null; int64 in the spatial file)
 frame: string
 img_w: int64
 img_h: int64
 type: string
 question: string
 answer: string
-answer_bbox: list<item: int64>
-evidence: null on general files;
-  on the spatial file: struct<
-    distances_2d_px: struct<<catalog object name>: double, ...>  (one entry
-      per published-catalog object name present in the scene, ~70 keys),
-    reference: string
-  >
+answer_bbox: list<int64>
+evidence: struct<distances_2d_px: struct<name: double, ...>, reference: string>
+                               (spatial file; one distance per catalog object)
 schema_version: string
 generator_version: string
 dataset_revision: string
 sample_id: string
 source_type: string
 source_question: string
-target_image: struct<member: string, repo_id: string, revision: string, shard: string>
-reference_image: struct<crop_bbox: list<item: int64>, member: string, repo_id: string, revision: string, shard: string>
+target_image: struct<member, repo_id, revision, shard: string>
+reference_image: struct<crop_bbox: list<int64>, member, repo_id, revision,
+                        shard: string>
 target_local_mask_id: int64
 target_source_catalog_id: string
 target_object_id: string
@@ -229,22 +130,52 @@ reference_source_catalog_id: string
 reference_object_id: string
 reference_global_object_id: int64
 reference_frame: string
-reference_mask_bbox: list<item: int64>
-reference_crop_bbox: list<item: int64>
+reference_mask_bbox: list<int64>
+reference_crop_bbox: list<int64>
 reference_crop_sha256: string
-attribute_kind: string on general files, null on spatial
-attribute_value: string on general files, null on spatial
-attribute_material: string on general files, null on spatial
-attribute_function: string on general files, null on spatial
+attribute_kind: string         (attribute files)
+attribute_value: string        (attribute files)
+attribute_material: string     (attribute files)
+attribute_function: string     (attribute files)
 answer_center_x_norm: double
 answer_center_y_norm: double
 answer_is_centered: bool
-different_category: bool on general files, null on spatial
-source_sample_id: string             (general files only)
-_dedup_key: string                   (spatial file only)
+different_category: bool       (attribute files)
+source_sample_id: string       (attribute files)
+_dedup_key: string             (spatial file; build bookkeeping)
 ```
 
-Note: unlike the published `spatial_bbox.parquet`'s `evidence` column
-(documented above as a JSON-encoded string), `incontext_mos_spatial_bbox.parquet`'s
-`evidence` is a native nested Parquet struct, not a JSON string -- read it
-directly with PyArrow/pandas rather than `json.loads`.
+</details>
+
+## Validation and provenance
+
+In-context image locators are pinned to dataset commit `93e31d378f1f98eca18a7aa01a2279c9f332440c`.
+The in-context generator,
+[`gt_incontext.py`](https://github.com/IRVLUTD/RPX/blob/main/data/mask_annotation/visual_grounding_gt/vqa_gt/gt_incontext.py),
+was released from commit `fbf2fd03cd9972d159ac954cfed8fd194deb8d87`.
+
+<details>
+<summary>What the validation report checks</summary>
+
+Almost every check in `VALIDATION_REPORT.json` runs on every row: box validity, identity and
+reference-visibility rules, duplicate sample IDs, semantic duplicates (raw and alias-normalized),
+reference crop availability and hashes, pinned revisions, row-for-row pairing with the published
+single-image files, and frame coverage across all MOS and Ego cells. Each of the 318,017 spatial rows
+equals one published `spatial_farthest` row.
+
+The one sampled check is an independent recomputation of the farthest-object geometry on a subset of
+scenes, run in addition to the exhaustive equality check above.
+
+</details>
+
+| File | SHA-256 |
+|:--|:--|
+| `incontext_mos_attribute_bbox.parquet` | `86a334247198848390c158537a235937dda2d386c022d56ede3bf135d798a94b` |
+| `incontext_ego_attribute_bbox.parquet` | `398565fc10b58f5beec3f237c5efb8cf0fd18123fed2782f6ab1c15561a9c186` |
+| `incontext_mos_spatial_bbox.parquet` | `6652cba3b471166e4f2dd366d9434495e63a7b80b5271cca61eeceb2ac0ec156` |
+
+---
+
+<sub>Part of <a href="https://github.com/IRVLUTD/RPX">RPX</a> ·
+<a href="https://irvlutd.github.io/RPX/toolkit-docs/">Toolkit docs</a> ·
+<a href="https://huggingface.co/datasets/IRVLUTD/RPX">Dataset card</a></sub>