Narendhiranv04/mock-scene79-gaming2-design2

Mock Scene79 Gaming2 Design-2 Full Release

This dataset is the full Design-2 release for scene79.gaming2. The important distinction is:

• the canonical RGB + depth + camera-pose universe is the full raw 753-frame scene
• the official segmentation-style annotations only exist for a verified 55-frame subset
• nvs is currently a scaffold config with zero rows because explicit source-target view groupings are not defined safely in the raw data

The goal of this release is to keep the repo structure clean and stable while being honest about what is and is not actually annotated.

At A Glance

• Scene id: scene79.gaming2
• Phase mapping: 0=cluttered, 1=declutter_transition, 2=uncluttered
• Canonical RGB source: full raw <phase>/rgb, republished under core/rgb/phase=...
• Full raw frame count: 753 = 251 + 251 + 251
• Depth coverage: full 753
• Camera pose coverage: full 753
• Segmentation / tracking / visual grounding coverage: 55 = 29 + 4 + 22
• Correspondence scaffolds: 753
• NVS scaffolds: 0

Why Some Tasks Are 55 Instead Of 753

This is coming from the raw source tree itself, not from filtering in the benchmark code.

In the raw dataset:

• 0/rgb, 1/rgb, 2/rgb: 251 frames each
• 0/depth, 1/depth, 2/depth: 251 frames each
• 0/cam_pose, 1/cam_pose, 2/cam_pose: 251 frames each

But segmentation-style assets only exist for the following verified subset:

• sam2/0/masks_verified: 29
• sam2/1/masks_verified: 4
• sam2/2/masks_verified: 22
• bboxes_verified/0: 29
• bboxes_verified/1: 4
• bboxes_verified/2: 22

The raw SAM2 masks also have the same coverage:

• sam2/0/masks: 29
• sam2/1/masks: 4
• sam2/2/masks: 22

So this repo does not pretend there are 753 segmentation masks when there are only 55 on disk. That is why:

• segmentation = 55
• tracking = 55
• visual_grounding = 55

Why NVS Is Empty

nvs is intentionally scaffold-only right now.

The raw scene contains RGB, fisheye, and pose assets, but it does not provide a trustworthy explicit definition of:

• target views
• source view sets
• novel-view synthesis evaluation pairs
• enough camera metadata to publish those pairings as an official benchmark task without inventing assumptions

So the release keeps the nvs config and manifest locations stable, but the actual tables are empty until real NVS pairings are defined.

Manifest Coverage

task	total_rows	cluttered	declutter_transition	uncluttered	status
master	753	251	251	251	full canonical sample universe
depth	753	251	251	251	full official GT depth coverage
camera_pose	753	251	251	251	full official pose coverage
segmentation	55	29	4	22	verified subset only
tracking	55	29	4	22	derived from verified segmentation subset
visual_grounding	55	29	4	22	segmentation-linked scaffold on verified subset
correspondences	753	251	251	251	multiview linkage scaffold over full set
nvs	0	0	0	0	scaffold only

Repo Organization

The primary release content lives under scene=scene79.gaming2/.

At the dataset-repo top level you may also see supporting docs and tools from earlier syncs. The core release to use is:

scene=scene79.gaming2/
├── metadata/
├── indexes/
├── core/
├── shared/
├── tasks/
└── extras/

Top-Level Files

• README.md: dataset card and structure guide
• BUILD_SUMMARY.md: how this release was built
• VALIDATION_REPORT.md: validation notes for the current release
• HF_UPLOAD_REPORT.md: upload summary for the current Hugging Face sync
• tools/: local fetch / validation / upload helpers

scene=scene79.gaming2/metadata/

• phase_map.json: maps raw phase ids to semantic names
• scene_info.json: scene-level summary, chosen RGB source, counts, notes

scene=scene79.gaming2/indexes/

These are the main parquet manifests.

• master.parquet: one canonical row per RGB frame in the release
• depth.parquet: depth benchmark manifest
• camera_pose.parquet: pose benchmark manifest
• segmentation.parquet: verified segmentation subset
• tracking.parquet: sequence-style manifest built on the verified segmentation subset
• visual_grounding.parquet: grounding scaffold built on the verified segmentation subset
• correspondences.parquet: multiview linkage scaffold using RGB + fisheye + pose assets
• nvs.parquet: empty scaffold manifest for future NVS work

scene=scene79.gaming2/indexes/hf/

These are Hugging Face split manifests. Each task gets:

• phase=cluttered.parquet
• phase=declutter_transition.parquet
• phase=uncluttered.parquet

This is what powers the Hub configs and splits in the dataset card.

scene=scene79.gaming2/core/

Shared image assets that should not be duplicated across tasks.

• core/rgb/phase=...: canonical RGB frames
• core/fisheye_left/phase=...: left fisheye views
• core/fisheye_right/phase=...: right fisheye views

scene=scene79.gaming2/shared/

Shared geometry used across multiple tasks.

• shared/geometry/cam_pose/phase=...: camera pose files

scene=scene79.gaming2/tasks/

Task-specific assets and eval lists.

• tasks/depth/gt/phase=...: official depth ground truth for all 753 frames
• tasks/segmentation/bboxes/phase=...: verified boxes for the 55 annotated frames
• tasks/segmentation/masks/phase=...: verified masks for the same 55 frames
• tasks/tracking/eval_lists/tracking.parquet: tracking eval manifest on the verified subset
• tasks/camera_pose/eval_lists/camera_pose.parquet: pose eval manifest
• tasks/visual_grounding/eval_lists/visual_grounding.parquet: grounding eval scaffold
• tasks/correspondences/eval_lists/correspondences.parquet: correspondence linkage scaffold
• tasks/nvs/eval_lists/nvs.parquet: empty NVS scaffold

scene=scene79.gaming2/extras/

Non-official or support assets.

• extras/weak_labels/bboxes_raw: raw boxes
• extras/weak_labels/bboxes_propagated: propagated boxes
• extras/weak_labels/masks_raw: raw SAM2 masks
• extras/visualizations/palette: palette images
• extras/visualizations/contour_masks: contour mask views
• extras/visualizations/rgb_and_mask: RGB+mask composites
• extras/provenance: provenance extracts and supporting metadata

Task Semantics

depth

• Full official task on all 753 frames
• Each row links canonical RGB to GT depth
• This is the manifest used by the RPX monocular depth benchmark

camera_pose

• Full official task on all 753 frames
• Each row links RGB, fisheye views, and camera pose

segmentation

• Official segmentation only where verified labels exist
• Coverage is exactly the verified 55-frame subset
• No synthetic expansion to 753 was performed

tracking

• Currently a sequence-style manifest over the same verified segmentation subset
• It reuses the verified bboxes/masks already present
• It does not claim separate dense tracking labels beyond that subset

visual_grounding

• Currently a spatial-grounding scaffold tied to verified segmentation regions
• Textual questions/answers are not available yet

correspondences

• This is currently a linkage scaffold, not dense pixel correspondence supervision
• It connects the canonical RGB frame with fisheye views and pose assets across all 753 frames

nvs

• Present for schema stability
• Empty until trustworthy NVS pairings are defined

How To Load The Dataset

Python datasets

from datasets import load_dataset

depth_ds = load_dataset(
    "Narendhiranv04/mock-scene79-gaming2-design2",
    name="depth",
    split="cluttered",
)

Hugging Face CLI

hf download Narendhiranv04/mock-scene79-gaming2-design2 \
  --repo-type dataset \
  --include "scene=scene79.gaming2/indexes/*" \
  --include "scene=scene79.gaming2/tasks/depth/*" \
  --local-dir ./mock_scene79_local

Benchmarking Note

For monocular depth estimation, use:

• scene=scene79.gaming2/indexes/depth.parquet

That manifest now covers the full 753 RGB + GT depth pairs.

Limitations

• Segmentation-style annotations are only available for 55 frames.
• tracking and visual_grounding therefore also remain 55-frame tasks.
• correspondences is a linkage scaffold, not dense ground-truth correspondences.
• nvs is scaffold-only and intentionally empty.
• Difficulty labels are represented but currently all rows are unassigned.

Future Expansion

To make this repo fully dense across tasks, the next missing pieces would be:

1. segmentation masks for the remaining 698 frames
2. verified boxes for those same frames
3. explicit grounding text labels
4. real correspondence labels
5. trustworthy NVS source-target pair definitions