Yukki1011/PromptDepth

MRQ Dataset

MRQ is a large-scale synthetic computer vision dataset with rendered multi-scene trajectories. It provides paired RGB frames, depth maps, optical-flow encodings, instance labels, camera poses, object/id mappings, and tracking annotations for depth, flow, segmentation, pose, and synthetic-to-real vision research.

Dataset Composition

The local source dataset is approximately 4.8 TB. The current Hub release is generated by the scene-packing upload pipeline in pack_and_upload_scenes.py and contains:

The Dataset Viewer contains 32 downsampled samples from Rome, DekoClass_night, TropicalRainForest, and ChemicalPlant for quick inspection of RGB, inverse-depth visualization, optical flow, instance labels, and camera metadata. These viewer samples are previews only; the full-resolution source remains in the scene archives.

Component	Count
Top-level archive entries	42
Trajectory folders (pathXX)	78
Scene config files	41
RGB frames (image/*.png)	197,877
Depth maps (depth/*.png)	197,577
Optical-flow maps (flow/*.png)	197,577
Scaled-flow variants (flow_scaled/*.png)	90,329
Instance label arrays (instance/*.npy)	197,577
Camera pose files (camera/*.json)	197,577
Tracking annotation files (track/*.txt)	407,325

Scene/archive entries include indoor, outdoor, urban, industrial, classroom, market, parking, forest, library, office, and stylized environments such as CityStreets, DeepMarket, ModernCity, TropicalRainForest, ParkingGarage, Rome, Vol8, and others.

Hub Release Layout

To make the dataset practical to upload and download from Hugging Face, the release is stored as one compressed archive per top-level scene entry:

archives/
├── AIUE_V02_002.tar.zst
├── BikeShop.tar.zst
├── CityStreets.tar.zst
├── DeepMarket.tar.zst
└── ...
archive_manifest.json
assets/
└── 195d399c3d6aa09243d813cd8d54d1ae.mp4
tools/
└── depth2pointcloud_downsampled.py

archive_manifest.json lists the archive entries included in the release. Each archive expands back to the original scene directory name:

tar --use-compress-program=zstd -xf archives/CityStreets.tar.zst

Scene Structure

Most scenes contain one or more trajectory folders named pathXX:

<scene>/
├── scene_config.json
└── pathXX/
    ├── camera/       # per-frame camera pose JSON files
    ├── depth/        # 16-bit grayscale PNG depth maps
    ├── flow/         # 16-bit RGB PNG optical-flow encodings
    ├── flow_scaled/  # optional generated flow variant
    ├── image/        # rendered RGB PNG frames
    ├── instance/     # NumPy instance label arrays
    ├── track/        # tracking correspondences/annotations
    ├── id2newid.json
    └── name2id.json

Some archives may contain auxiliary visualization or generated-output folders in addition to canonical pathXX trajectories.

Upload Processing Notes

The scene-packing upload logic stages each scene before archiving:

• RGB/image PNG files are resized to fit within 1280 x 720 using Lanczos interpolation.
• Depth and optical-flow PNG encodings are resized to fit within 1280 x 720 using Triangle interpolation.
• Instance .npy label maps are resized with nearest-neighbor sampling to preserve ids.
• Camera JSON files are preserved as pose files. If downstream code uses intrinsics, scale fx/cx by resized_width/original_width and fy/cy by resized_height/original_height.
• Each staged scene includes _resize_metadata.json documenting resized files and modality-specific interpolation.
• Scene archives are processed in ascending source-directory size order, so smaller scenes are packed and uploaded before larger scenes.

This release format avoids millions of small Hub files while preserving the original per-scene/per-trajectory organization after extraction.

Downsampled Visualization Tools

The repository includes tools/depth2pointcloud_downsampled.py, a point-cloud utility adapted from the verified local depth-to-pointcloud script. It first resizes RGB/depth to the Hub upload resolution, keeps camera extrinsics unchanged, recomputes FOV-based intrinsics from the downsampled image size, and then fuses colored point clouds.

For the current downsampled point-cloud visualization convention, the camera vertical axis is flipped relative to the original helper:

z_cam = (v_grid - half_rows) * tan_fov_y * depths

Example:

python tools/depth2pointcloud_downsampled.py \
  --path-dir MRQ/DekoClass_night/path00 \
  --start 0 \
  --frames 80 \
  --subsample-step 6 \
  --camera-fov 90 \
  --output outputs/DekoClass_night_downsampled_pointcloud.ply

Demo Videos

The following Rome demo was generated from Rome/path00, frames 000000 through 000299. The original frames are 1920 x 1080 and are resized to 1280 x 720 before visualization.

File Formats

• RGB images: PNG, 8-bit RGB.
• Depth maps: PNG, 16-bit grayscale.
• Optical flow: PNG, 16-bit RGB normalized encoding.
• Instance labels: NumPy .npy arrays.
• Camera poses: JSON files with position and orientation.
• Tracks: .txt files.
• Scene configuration and id maps: JSON files.

Example camera pose file:

{
  "position": [-105.0, -93.2143783569336, 20.6917781829834],
  "orientation": [
    0.9553737342265323,
    0.03291562659926051,
    -0.1167528882766679,
    0.2693443011364547
  ]
}

Loading Example

from pathlib import Path
import json
import numpy as np
from PIL import Image

root = Path("MRQ")
scene = "CityStreets"
path = "path00"
frame = "000479"

rgb = Image.open(root / scene / path / "image" / f"{frame}.png")
depth = Image.open(root / scene / path / "depth" / f"{frame}.png")
flow = Image.open(root / scene / path / "flow" / f"{frame}.png")
instance = np.load(root / scene / path / "instance" / f"{frame}.npy")

with open(root / scene / path / "camera" / f"{frame}.json", "r") as f:
    camera = json.load(f)

Intended Use

MRQ is intended for research and development in monocular and multi-frame depth estimation, optical flow, visual odometry/camera pose, instance segmentation, object tracking, and synthetic-to-real computer vision experiments.

Notes

• The dataset is large. Prefer downloading only the scene archives needed for an experiment.
• Generated helper scripts, local editor folders, upload caches, and intermediate upload state are not part of the intended dataset content.
• Please verify licensing and redistribution terms for downstream public use.

Citation

If you use this dataset, please cite:

@inproceedings{wang2026promptdepth,
  title={PromptDepth: Efficient and Promptable Geometric 3D Vision Model for Embodied Intelligence},
  author={Wang, Xianyun and Miao, Jiaxu and Xu, Tian and Wang, Siyuan and Li, Yuehao and Hu, Haoyang and Xiao, Jun and Tian, Yonghong and Yu, Jun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={28074--28085},
  year={2026}
}