Add dataset card with full documentation

README.md

@@ -0,0 +1,168 @@
+---
+license: cc-by-nc-4.0
+task_categories:
+  - robotics
+  - video-classification
+  - depth-estimation
+tags:
+  - egocentric
+  - rgbd
+  - manipulation
+  - mcap
+  - ros2
+  - imu
+  - pointcloud
+  - depth
+  - kitchen
+  - household
+  - slam
+  - hand-tracking
+  - body-pose
+pretty_name: "MCAP-Housing: Egocentric RGB-D Manipulation Dataset"
+size_categories:
+  - 1K<n<10K
+---
+
+# MCAP-Housing: Egocentric RGB-D Household Manipulation Dataset
+
+**MCAP-Housing** is an egocentric **RGB + Depth + IMU** dataset of human household manipulation activities, captured with iPhone LiDAR and packaged in robotics-native `.mcap` (ROS2) format. Designed for robotics research, policy learning, and embodied AI.
+
+---
+
+## Quick Facts
+
+| Property | Value |
+|----------|-------|
+| **Modalities** | Synchronized RGB + 16-bit Depth + IMU + Point Clouds |
+| **Resolution (RGB)** | 1920 × 1440 @ 60 FPS |
+| **Depth** | 16-bit millimeter, LiDAR-sourced (256×192 native, aligned to RGB) |
+| **Point Clouds** | Per-frame colored XYZRGB (up to 50k points) |
+| **IMU** | 6-axis (accel + gyro) + magnetometer + gravity + orientation @ 60 Hz |
+| **Pose** | 6DoF camera pose (world → camera transform) per frame |
+| **Activities** | 10 household manipulation sequences |
+| **Total Frames** | ~30,000 synchronized RGB-D pairs |
+| **Total Size** | ~30 GB |
+| **Container** | `.mcap` with ROS2 CDR serialization |
+| **Compression** | zstd (~95% ratio) |
+
+---
+
+## What's Included Per Sequence
+
+Each `.mcap` file contains **11 synchronized ROS2 topics**:
+
+| Topic | Message Type | Description |
+|-------|-------------|-------------|
+| `/camera/rgb/compressed` | `sensor_msgs/CompressedImage` | JPEG-encoded RGB frames |
+| `/camera/depth/aligned` | `sensor_msgs/Image` | Raw 16-bit depth aligned to RGB |
+| `/camera/depth/filtered` | `sensor_msgs/Image` | Bilateral-filtered depth (hole-filled) |
+| `/camera/depth/colorized` | `sensor_msgs/Image` | Turbo-colormap depth visualization |
+| `/camera/points` | `sensor_msgs/PointCloud2` | Colored XYZRGB point cloud |
+| `/camera/camera_info` | `sensor_msgs/CameraInfo` | Per-frame intrinsics (fx, fy, cx, cy) |
+| `/tf` | `tf2_msgs/TFMessage` | 6DoF camera pose (world → camera) |
+| `/imu` | `sensor_msgs/Imu` | Linear acceleration + angular velocity |
+| `/imu/gravity` | `geometry_msgs/Vector3Stamped` | Gravity vector |
+| `/imu/orientation` | `geometry_msgs/QuaternionStamped` | Device orientation quaternion |
+| `/imu/mag` | `sensor_msgs/MagneticField` | Magnetometer readings |
+
+---
+
+## Sequences
+
+| Sequence | Duration | Frames | Size |
+|----------|----------|--------|------|
+| Chopping | 58s | 3,449 | 3.2 GB |
+| Filling water bottles | 57s | 3,442 | 3.5 GB |
+| Folding T shirts | 66s | 3,977 | 3.7 GB |
+| Grating in the Kitchen | 59s | 3,538 | 3.1 GB |
+| Lighting a stove | 25s | 1,507 | 1.4 GB |
+| Peeling Banana | 48s | 2,854 | 2.5 GB |
+| Preparing a solution | 42s | 2,539 | 2.4 GB |
+| Sweeping | 32s | 1,904 | 2.3 GB |
+| Transferring processed spices | 39s | 2,368 | 2.2 GB |
+| Working mortar and Pestle | 112s | 6,701 | 6.3 GB |
+
+---
+
+## Optional Derived Signals (Available on Request)
+
+Beyond the raw synchronized streams, the following derived signals can be provided:
+
+- **Ego-motion / trajectories** (VIO-style) — smooth, drift-corrected camera trajectories
+- **SLAM reconstructions** — dense maps, optimized trajectories, keyframe selection
+- **Accurate body pose estimation** — full skeletal tracking during manipulation
+- **State-of-the-art 3D hand landmarks** — true 3D hand joint positions, not 2D reprojections
+- **Additional QA layers and consistency checks** tailored to your training setup upon request
+
+---
+
+## Data Quality
+
+- Tight RGB ↔ Depth ↔ IMU synchronization (all streams at 60 Hz)
+- Per-frame camera intrinsics (not a single fixed calibration)
+- Per-frame 6DoF pose from ARKit visual-inertial odometry
+- Depth hole-filling and bilateral filtering provided as separate topics
+- zstd compression with ~95% ratio for efficient storage and transfer
+
+---
+
+## Getting Started
+
+### Inspect a file
+```bash
+pip install mcap
+
+mcap info Chopping.mcap
+```
+
+### Read in Python
+```python
+from mcap.reader import make_reader
+from mcap_ros2.decoder import DecoderFactory
+
+with open("Chopping.mcap", "rb") as f:
+    reader = make_reader(f, decoder_factories=[DecoderFactory()])
+    for schema, channel, message, decoded in reader.iter_decoded_messages():
+        if channel.topic == "/camera/rgb/compressed":
+            print(f"RGB frame at t={message.log_time}, size={len(decoded.data)} bytes")
+        elif channel.topic == "/camera/depth/aligned":
+            print(f"Depth frame: {decoded.width}x{decoded.height}, encoding={decoded.encoding}")
+        elif channel.topic == "/camera/points":
+            print(f"Point cloud: {decoded.width} points")
+```
+
+### Visualize
+Open any `.mcap` file directly in [Foxglove Studio](https://foxglove.dev/) for full 3D visualization of RGB, depth, point clouds, and transforms.
+
+### Dependencies
+```bash
+pip install mcap mcap-ros2-support numpy opencv-python
+```
+
+---
+
+## Intended Uses
+
+- **Policy and skill learning** — imitation learning, VLA pre-training
+- **Action detection and segmentation** — temporal activity recognition
+- **Hand and body pose estimation** — grasp analysis, manipulation understanding
+- **Depth-based reconstruction** — SLAM, scene understanding, 3D mapping
+- **World-model training** — ego-motion prediction, scene dynamics
+- **Sensor fusion research** — RGB-D-IMU alignment and calibration
+
+---
+
+## License
+
+This dataset is released under **CC-BY-NC-4.0**. Free for research and non-commercial use with attribution. For commercial licensing, contact us.
+
+**Required attribution:** *"This work uses the MCAP-Housing dataset (Cortex Data Labs, 2025)."*
+
+---
+
+## Contact
+
+- **Email:** shashin.bhaskar@gmail.com
+- **Organization:** [Cortex Data Labs](https://huggingface.co/cortexdatalabs)
+
+For custom data capture, derived signals (hand tracking, SLAM, body pose), or commercial licensing, reach out directly.