min99ian/realsense-l515-rgbd-raw

RealSense L515 RGB-D Raw

Single-frame raw RGB-D captures from an Intel RealSense L515. The directory layout and field contract align with the Isaac Sim raw-frame output of isaac_sim_data_collector, so frames can be back-projected and converted to PTv3 / Pointcept datasets with existing tooling.

This repository contains raw data only (no converted point clouds / PTv3 splits).

Download

pip install huggingface_hub
hf download min99ian/realsense-l515-rgbd-raw --repo-type dataset --local-dir ./realsense_raw

Or:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="min99ian/realsense-l515-rgbd-raw",
    repo_type="dataset",
    local_dir="./realsense_raw",
)

Directory layout

realsense_raw/
  scene_xxx/
    capture_metadata.json
    calibration.json
    frames/
      frame_XXXX/
        rgb.npy / rgb.png
        depth.npy
        normals.npy
        camera_params.json
        pose.json
        frame_metadata.json
        depth_raw_uint16.npy / .png
        depth_m_float32.npy
        depth_valid_mask.npy
        color_original.png
        infrared.png          # optional

Per-frame files

File	Content
rgb.npy / rgb.png	uint8 H×W×3, RGB, aligned to depth
depth.npy	float32 meters, distance_to_image_plane; invalid = inf
normals.npy	float32 H×W×3, model camera frame (X right, Y up, Z backward); invalid = 0
camera_params.json	Intrinsics and related fields (Isaac field names + realsense extras)
pose.json	Identity placeholder; pose_available=false
frame_metadata.json	RealSense-specific metadata
depth_raw_uint16.*	Raw Z16 depth
color_original.png	Unaligned original color (e.g. 1920×1080)
depth_m_float32.npy	depth_raw * depth_scale (invalid = 0)
depth_valid_mask.npy	depth_raw > 0

Scene-level:

• capture_metadata.json: device info, stream profiles, frame list
• calibration.json: depth/color intrinsics & distortion, depth→color extrinsics, depth scale

Capture setup (reference)

• Device: Intel RealSense L515
• Depth: 1024×768 @ 30 FPS, Z16
• Color: 1920×1080 @ 30 FPS, RGB8
• IR: 1024×768 @ 30 FPS, Y8 (optional)
• Depth range filter at capture: about 0.25–9.0 m
• No semantic GT, no real world pose

Known limitations

• Normals are estimated from the depth point cloud (not rendered mesh normals); only camera-facing orientation semantics are guaranteed.
• pose.json is an identity placeholder; use the camera coordinate frame.
• L515 depth FOV is larger than color FOV, so some depth pixels have no real color.
• Real scenes have no labels; downstream conversion must allow missing labels.

Convert to PTv3

Capture and conversion code lives in the companion repo realsense_data_collector:

python tools/convert_realsense_to_ptv3.py `
    --input-root <downloaded>/realsense_raw `
    --output-root data/realsense_ptv3 `
    --split test `
    --bags scene_real_00 `
    --min-depth 0.25 --max-depth 9.0 `
    --max-points 120000

Citation

If you use this dataset, please credit the source and device model (Intel RealSense L515) and link to this Hugging Face dataset page.

License

CC BY 4.0
(Change to CC-BY-NC-4.0 or another license before upload if needed, and update the YAML license field accordingly.)

---

RealSense L515 RGB-D Raw（中文）

Intel RealSense L515 采集的单帧原始 RGB-D 数据。目录结构与字段契约对齐 isaac_sim_data_collector 的 Isaac Sim 原始帧输出，便于后续反投影与 PTv3 / Pointcept 转换。

本仓库只包含 raw（不含转换后的点云 / PTv3 数据集）。

下载

pip install huggingface_hub
hf download min99ian/realsense-l515-rgbd-raw --repo-type dataset --local-dir ./realsense_raw

或：

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="min99ian/realsense-l515-rgbd-raw",
    repo_type="dataset",
    local_dir="./realsense_raw",
)

目录结构

realsense_raw/
  scene_xxx/
    capture_metadata.json
    calibration.json
    frames/
      frame_XXXX/
        rgb.npy / rgb.png
        depth.npy
        normals.npy
        camera_params.json
        pose.json
        frame_metadata.json
        depth_raw_uint16.npy / .png
        depth_m_float32.npy
        depth_valid_mask.npy
        color_original.png
        infrared.png          # optional

每帧字段

文件	内容
rgb.npy / rgb.png	uint8 H×W×3，RGB，已对齐到 Depth
depth.npy	float32，米，distance_to_image_plane；无效为 inf
normals.npy	float32 H×W×3，model camera frame（X 右、Y 上、Z 向后）；无效为 0
camera_params.json	相机内参等（Isaac 字段名 + realsense 附加字段）
pose.json	identity 占位；pose_available=false
frame_metadata.json	RealSense 专属元数据
depth_raw_uint16.*	原始 Z16
color_original.png	未对齐原始彩色图（如 1920×1080）
depth_m_float32.npy	depth_raw * depth_scale（无效为 0）
depth_valid_mask.npy	depth_raw > 0

场景级：

• capture_metadata.json：设备、流配置、帧列表
• calibration.json：depth/color 内参与畸变、depth→color 外参、depth scale

采集配置（参考）

• 设备：Intel RealSense L515
• Depth：1024×768 @ 30 FPS, Z16
• Color：1920×1080 @ 30 FPS, RGB8
• IR：1024×768 @ 30 FPS, Y8（可选）
• 深度范围过滤（采集时）：约 0.25–9.0 m
• 无语义 GT、无真实 world pose

已知限制

• 法线由深度点云估计（非渲染法线），只保证朝向相机的方向语义。
• pose.json 为占位 identity，camera 坐标系下使用即可。
• L515 深度 FOV 大于彩色 FOV，部分深度像素无真实颜色。
• 真实场景无标签；下游转换需允许 missing labels。

转换到 PTv3

采集与转换代码见配套仓库 realsense_data_collector：

python tools/convert_realsense_to_ptv3.py `
    --input-root <downloaded>/realsense_raw `
    --output-root data/realsense_ptv3 `
    --split test `
    --bags scene_real_00 `
    --min-depth 0.25 --max-depth 9.0 `
    --max-points 120000

引用

若使用本数据，请注明数据来源与设备型号（Intel RealSense L515），并链接本 Hugging Face 数据集页面。

许可证

CC BY 4.0
（上传前可按需要改为 CC-BY-NC-4.0 等，并同步修改本文件 YAML license 字段。）