tracking/metadata/text_initialization_v1/README.md

Tracking text prompts (v1)

Fixed text prompts, such as black hammer or silver alarm clock, for every tracked object in RPX. In the paper they initialize the text-promptable trackers (Grounded SAM 2 and SAM 3.1) in place of a first-frame mask or box.

There are 70 objects, each with one prompt of the form <primary color> <object name>. Prompts are given per clip, and there are 400 clips: 100 scenes × 3 phases (MOS) plus 100 egocentric recordings (Ego). That makes 2,739 rows, one per object per clip. A prompt stays the same for the whole clip, including frames where the object is occluded. Rows join to the released masks on scene_id, kind, phase and mask_index.

Usage · Files · How prompts are built · Known limitations · Reproducing

Usage

import pandas as pd

vocab = pd.read_parquet("scene_condition_vocab.parquet")
clip = vocab.query("scene_id == 'scene001' and kind == 'mos' and phase == 1")

# {1: 'tan lion', 2: 'white boot', ...}
prompts = dict(zip(clip.mask_index, clip.prompt_text))

phase is 0, 1 or 2 (Clutter, Interaction, Clean) for MOS and empty for Ego. The .jsonl files hold the same rows as the Parquet files.

Files

File	Rows	Contents
scene_condition_vocab.parquet / .jsonl	2,739	One row per object per clip
object_catalog_vocab.parquet / .jsonl	70	One row per object
manifest_summary.json		Counts, sources, validation, hashes
SHA256SUMS		SHA-256 of every file here

Clip rows hold mask_index, prompt_text, the object's identity and the annotation it was joined from. Object rows hold the name, all authored colors, materials, functions and categories, primary_color and prompt_text. manifest_summary.json also lists the known limitations below.

How prompts are built

prompt_text is primary_color + " " + display_name.

display_name is the object's catalog name with underscores replaced by spaces (air_duster_can becomes air duster can). primary_color is the first color in the object's authored questionnaire, lowercased, with gray spelled grey. It is fixed per object, not the color that happens to be most visible in a given clip.

Object identities come from the dataset's object catalog and each clip's released mask_to_object.json. The VQA attribute table is only used as a cross-check.

Known limitations

• 60 Ego mask entries are labeled unknown/unlabeled in the released annotations. They have no identity, so they get no prompt. They are listed in manifest_summary.json.
• 252 of 658,254 visible Ego object occurrences (0.038%) use a mask index that has no identity at all in the released annotations. They fall in 14 of the 100 Ego clips, and 179 of them are index 8 in scene058. MOS is not affected. A tracker may therefore see a mask index in those clips with no matching row.
• Colors are free text, so a primary_color can be black and white. An object has the same prompt in every scene it appears in.

Reproducing

cd benchmark/scripts/tracking_text_vocab
python3 build_tracking_text_vocab.py --out-dir /path/to/out
python3 validate_tracking_text_vocab.py --bundle-dir /path/to/out \
    --report /path/to/out/validation_report.json

The validator exits nonzero on this bundle because of the 252 occurrences above; every other check passes. To confirm your run matches the release, compare its missing_mappings_count with the one in manifest_summary.json. The full coverage check streams every mask archive one clip at a time and deletes each when done, so disk use stays bounded. --skip-exhaustive-masks skips it for quick iteration; don't use it for a release check.

Design notes

Why not use the VQA color questions

vqa/attribute.parquet only asks an attr_single_color question when color alone identifies one visible object in that frame. It is a discrimination question, not a list of attributes. Building prompts from it would drop every object whose color is shared with something else on screen, or that was never the subject of a color question. Its answers also use questionnaire synonyms ("cleaner", "male toy", "analog alarm clock", "joystick") rather than the catalog name.

Prompts are built instead from each object's catalog identity (object_name in manifest/object_catalog_v1.json) and its authored questionnaire (the color, material, function and category lists in objects_meta/<id>/questionnaire.json). Both exist for all 70 published objects. vqa/attribute.parquet is read only to cross-check the result and never fills in a missing attribute.

Source revisions

Each input is read at a fixed commit of the dataset:

Purpose	Revision
Catalog and questionnaires	93e31d378f1f98eca18a7aa01a2279c9f332440c
MOS sam2_meta and masks tars	2e2a387f7f93e98c177b2e039c141eacda94e5fc
Ego sam2_meta and masks tars	f082723002bad5800dd85e583115b4ea05734d31
VQA cross-check	vqa branch

The catalog revision is the dataset's main branch when the bundle was built. The two tracking revisions predate the color, material and function questions; their catalog copies only have category and name for all 70 objects, so they can't be the attribute source. The MOS revision is PINNED_DATASET_REVISION in benchmark/scripts/run_tracking.py on this repo's docker/tracking-all-yoloe branch, the revision the tracking runner pins for MOS; the Ego revision is PINNED_EGO_DATASET_REVISION in the same file. The VQA cross-check reads vqa/attribute.parquet only.

Join

1. manifest/object_catalog_v1.json lists the 70 published objects. Each entry's source_catalog_id is the join key. It is a string (for example "18.2") and is never converted to a number.
2. Each entry's questionnaire_path points to objects_meta/<id>/questionnaire.json. Its questions are matched against fixed patterns: "made of" gives material, "used for" function, "color" color, "category" category and "name" name. This is the same parsing as data/mask_annotation/visual_grounding_gt/vqa_gt/sos_catalog.py, copied rather than imported so that this package has no dependency on the annotation code.
3. For each clip, sam2/mask_to_object.json inside scenes/<scene>/<phase>/labels/sam2_meta/v1.tar (MOS) or scenes/<scene>/ego/labels/sam2_meta/v1.tar (Ego) maps each mask_index to {"id": <source_catalog_id>, "name": <name in that scene>}. Joining id to the catalog gives every tracked object in the clip its catalog name and attributes.
4. As a check, the object ids in vqa/attribute.parquet (non-null target_oid and the JSON-encoded answer_oids) are matched against the same catalog.

Normalization

• colors_raw, materials_raw, functions_raw and categories_raw keep the authored lists verbatim and in order.
• colors_normalized is colors_raw lowercased and stripped, with internal whitespace collapsed and gray written as grey, the same rule as _canon_color in gt_incontext.py.
• primary_color is always the first entry of colors_normalized. The full lists are kept next to it.
• display_name is canonical_name lowercased with underscores replaced by spaces. None of the 70 object_name values carries an instance suffix (suffixes live only in object_id and source_catalog_id, e.g. object_id="baking_tray.1" with object_name="baking_tray"), and all 70 are distinct, so no suffix stripping is needed.
• prompt_text is f"{primary_color} {display_name}".

unknown/unlabeled entries

Some Ego mask_to_object.json files contain {"id": "unknown", "name": "unlabeled"} for a mask that was never given an identity. Across all 400 clips this happens for 60 of 2,799 entries, all of them Ego and all with exactly this pair; MOS has none, and there are no other unresolved ids. With no name or color there is nothing to build a prompt from, so these entries are left out of scene_condition_vocab and listed in manifest_summary.json and the validator's report.

Ego mask indices with no identity

The validator checks every mask in every one of the 98,124 MOS and Ego frames. It found 252 visible object occurrences (0.038% of 658,254) whose mask index has pixels in labels/masks/v1.tar but no entry in that clip's mask_to_object.json, not even the placeholder above. MOS is clean (0 of 75,000 frames).

This is a gap in the published Ego annotations rather than in this pipeline. For example, scene058's mask_to_object.json lists indices 1, 3–7 and 9 but not 8. Ego clips also have no masks_aux tar, which MOS uses as a second identity source, so there is nothing to build a prompt from without inventing an identity.

14 of the 100 Ego clips are affected. 179 of the 252 occurrences are index 8 in scene058, which is present in 179 of that clip's 182 frames. The other 73 are one- to three-frame blips across 13 scenes. Several of those use indices in the range 10001–10009, which otherwise never occur in those clips, so they look like brief segmentation artifacts rather than missed objects. The full list of scenes, frames and indices is in manifest_summary.json under coverage_and_validation.exhaustive_mask_coverage_check.missing_mappings and in the validator's report.

The bundle covers every mask index that has an identity: all of MOS and 99.96% of Ego occurrences.

One prompt per clip, not per frame

scene_condition_vocab has one row per object per scene, kind and phase. A promptable tracker is usually initialized once at the start of a clip and needs a prompt for every object it may have to follow, including one that is occluded for a while and comes back. The per-frame mask check in validate_tracking_text_vocab.py only proves that every mask index appearing in any frame is covered by these clip-level rows. The bundle does not provide per-frame prompts.

Authored colors

• Colors are free-text questionnaire answers, not a fixed palette. Values such as black and white, grey and yellow or white black red and blue are real answers and appear verbatim in primary_color and prompt_text when they come first.
• primary_color is the first authored color. It is not necessarily the dominant or most visible color under a particular clip's lighting.
• Attributes belong to the object, not to a scene, so the same physical object has the same prompt everywhere it appears.

---

_{Part of RPX · Toolkit docs · Dataset}