---
language:
  - en
  - zh
pretty_name: R4D-Bench
tags:
  - referring-expression-segmentation
  - video-understanding
  - spatio-temporal
  - dynamic-scenes
  - instance-segmentation
task_categories:
  - image-segmentation
size_categories:
  - 10K<n<100K
license: other
dataset_summary: "Spatio-temporal referring segmentation on dynamic 4D scenes: 89 bilingual queries with dense mask GT across 12 scenes, plus Mask IoU and track-ID evaluation scripts."
---

# R4D-Bench

## What this dataset is

**R4D-Bench** is a benchmark for **spatio-temporal referring segmentation** in **dynamic (4D) scenes**: given a natural-language query, the model must return the correct **instance mask(s) over time**.

This release ships:

| Shipped | Count / note |
|---------|----------------|
| Scenes | **12** under `data/scenes/` |
| Referring queries | **89** (bilingual `text_en` / `text_zh`) |
| Dense mask GT | Per-query, per-frame polygons/RLE + PNG unions under `query_masks/` |
| Eval scripts | Mask IoU (`evaluate_mask_temporal.py`) and track-ID F1 (`evaluate_r4dgs.py`) |
| Optional subset | **36** queries (3 / scene) for quick debugging |

**Not required** for the shipped metrics: COLMAP, cameras, Segment-then-Splat outputs, or a full multi-view training tree. Original **Neu3D / HyperNeRF** imagery (when you need more RGB beyond what is in each scene folder) stays under **upstream licenses**.

> **Hugging Face Dataset Viewer:** this repo is a **folder-layout benchmark** (JSON + PNG masks), not a tabular Parquet split. The Hub “Dataset Viewer / test split” panel will stay unavailable — that is expected. Use **Files and versions** and the commands below.

---

## Quick start (3 steps)

```bash
# 1) Download this dataset repo (Hub UI, git lfs, or huggingface_hub snapshot_download)

# 2) Sanity-check mask metrics (predictions = GT → perfect scores)
python evaluation/evaluate_mask_temporal.py --self-check \
  --ground-truth scripts/new_predictions_ground_truth_all_queries.json \
  --queries-meta evaluation/R4D-Bench_queries.json

# 3) Track-ID metrics (GT track sets as a demo)
python evaluation/evaluate_r4dgs.py \
  --queries evaluation/R4D-Bench_queries.json \
  --predictions evaluation/R4D-Bench_predictions.json \
  --output evaluation/track_id_eval_report.json
```

**Dependencies:** Python 3.10+; `numpy`, `Pillow` for mask eval. Evaluation scripts live next to the data — run them from the **repo root**.

---

## Which files to open first

| Goal | Open this |
|------|-----------|
| Read all 89 queries (text + target tracks) | `evaluation/R4D-Bench_queries.json` |
| Track-ID ground truth map | `evaluation/R4D-Bench_predictions.json` (`query_id` → `target_track_ids`) |
| Dense mask ground truth (canonical, 89) | `scripts/new_predictions_ground_truth_all_queries.json` |
| Same GT + PNG path fields | `scripts/new_predictions_ground_truth_all_queries_with_paths.json` |
| Per-scene query source of truth | `data/scenes/<scene>/*_queries.json` |
| Per-frame answer PNGs | `data/scenes/<scene>/query_masks/frame_XXXXXX/<query_id>/…` |
| Track id → category cheat sheet | `data/track_metadata.csv` (docs only; **not** loaded by eval) |
| Path checklist | `DATASET_LAYOUT.md` |

**Optional 36-query subset:** `scripts/new_predictions_ground_truth_final.json` (+ `*_with_paths.json`). Same format, fewer queries — **not** a separate release tier.

---

## Scenes

Folders under `data/scenes/` (underscores):

`americano`, `coffee_martini`, `cook_spinach`, `cut_lemon`, `cut_roasted_beef`, `espresso`, `flame_salmon`, `flame_steak`, `keyboard`, `sear_steak`, `split_cookie`, `torchchocolate`

Each scene typically contains:

- `*_queries.json` — queries for that scene  
- scene track JSON (e.g. `americano.json`; `cook_spinach` uses `cook-spinach.json`)  
- COCO annotation JSON (`_annotations*.coco.json`)  
- `images/` — RGB frames used with the annotations  
- `query_masks/` — query-level union masks for the **89**-query set  
- `query_masks_36/` — masks for the optional **36**-query subset (when present)  
- `README.dataset.txt` / `README.roboflow.txt` — upstream provenance  

**Naming:** folder names use `_`; some JSON / `query_id` prefixes use `-` (e.g. `cook-spinach`, `split-cookie`).

---

## Task definition & conventions

- **Input:** referring expression (`text_en` / `text_zh`) + scene frames.  
- **Output (mask track):** per-`frame_id` foreground mask(s) for the referred target(s).  
- **Output (track-ID track):** set of `target_track_ids` (no pixels).  
- **Time index:** integer `frame_id` in GT / `existence_frames`, matching `query_masks/frame_XXXXXX/`.  
- **Left / right:** means **image / viewer** coordinates (left or right side of the picture), **not** the person’s anatomical left/right hand.  
- **Query types:** `A` / `B` / `C` in metadata (see fields below); many items need **motion** or **before / while / after** reasoning; some are multi-target or distractor-style.

### Query schema (`evaluation/R4D-Bench_queries.json`)

Each entry includes among others:

- `query_id`, `scene`, `text_en`, `text_zh`  
- `target_track_ids`, `target_count`  
- `query_type`, `requires_motion_understanding`  
- Optional: `temporal_relation`, `expected_temporal_anchor`, `distractor`, `note`

Dense GT entries in `scripts/new_predictions_ground_truth_all_queries.json` add `ground_truth` with `target_tracks`, `existence_frames`, and `frames[].masks[]` (`segmentation`, optional `mask_image`). Geometry in `segmentation` / PNGs is authoritative.

---

## How to evaluate your model

### 1) Mask-level (mIoU, mAcc, temporal Acc, vIoU, …)

```bash
python evaluation/evaluate_mask_temporal.py \
  --predictions path/to/your_predictions.json \
  --ground-truth scripts/new_predictions_ground_truth_all_queries.json \
  --output evaluation/mask_eval_report.json \
  --queries-meta evaluation/R4D-Bench_queries.json
```

**Prediction JSON** (either shape):

- `{ "query_id": { "frames": { "1": "path/to/mask.png", ... } } }`  
- or a list of `{ "frame_id", "mask_path" }` per query  

Paths resolve relative to `--repo-root` unless absolute. Useful flags: `--only-query-prefix`, `--only-queries`, `--debug-query`, `--iou-threshold`.

### 2) Track-ID (precision / recall / F1 on ID sets)

```bash
python evaluation/evaluate_r4dgs.py \
  --queries evaluation/R4D-Bench_queries.json \
  --predictions path/to/your_track_predictions.json \
  --output evaluation/track_id_eval_report.json
```

Your prediction file should use the same `query_id` keys as `evaluation/R4D-Bench_predictions.json` (each value = predicted track-id list/set).

---

## Repository layout

```
R4D-Bench/
├── DATASET_LAYOUT.md
├── README.md
├── evaluation/
│   ├── evaluate_mask_temporal.py
│   ├── evaluate_r4dgs.py
│   ├── R4D-Bench_queries.json      # 89 queries (eval metadata)
│   └── R4D-Bench_predictions.json  # query_id → target_track_ids (GT)
├── scripts/
│   ├── coco_scene_paths.py
│   ├── generate_instance_masks.py
│   ├── predictions_ground_truth.py
│   ├── enrich_ground_truth_with_mask_images.py
│   ├── new_predictions_ground_truth_all_queries.json
│   ├── new_predictions_ground_truth_all_queries_with_paths.json
│   ├── new_predictions_ground_truth_final.json              # optional 36
│   └── new_predictions_ground_truth_final_with_paths.json
└── data/
    ├── scenes/<scene>/    # images, COCO, tracks, queries, query_masks/
    ├── all_instance_masks/  # optional regeneratable snapshot
    └── track_metadata.csv   # optional human-readable reference
```

`upload_large.py` is a maintainer helper for Hub uploads; end users can ignore it.

---

## Regenerating masks / GT (maintainers)

```bash
# Per-instance PNGs from COCO (not the same as query union masks)
python scripts/generate_instance_masks.py --overwrite

# Rebuild 89-query dense GT from per-scene *_queries.json
python scripts/predictions_ground_truth.py --all-queries \
  --output scripts/new_predictions_ground_truth_all_queries.json

# Add mask_image / combined_mask_image path fields into JSON
python scripts/enrich_ground_truth_with_mask_images.py
```

---

## Relation to Segment-then-Splat & upstream data

- **Segment-then-Splat** is a separate pipeline. Typical StS trees (`sparse/`, merged multiview masks, PLY, …) are **not** part of this minimal annotation release.  
- Pair with **HyperNeRF / Neu3D** (or compatible) downloads under their original licenses when you need the full original capture.  
- Roboflow URLs / licenses: see each scene’s `README.dataset.txt` / `README.roboflow.txt`.

---

## Citation

If you use **R4D-Bench**, cite **this dataset repository** and the **original scene datasets** (HyperNeRF, Neu3D, Roboflow sources as applicable). Add BibTeX here after publication if desired.

---

## License

- **Annotations and code** in this repository: treat as **`license: other`** until a top-level `LICENSE` is added; contact maintainers if you need redistributable terms in writing.  
- **Scene imagery / upstream assets** remain under **Neu3D / HyperNeRF / Roboflow** (etc.) original terms. Do not redistribute raw imagery unless those licenses allow it.

---

## Contact

Questions on benchmark definition, evaluation protocol, or file formats: open a discussion/issue on this Hub dataset page or contact the maintainers.