Add Real4D metadata: docs/DATA_FORMAT.md

docs/DATA_FORMAT.md

@@ -0,0 +1,76 @@
+# Real4D data format
+
+## Paths
+
+The first path component identifies the source subset (`dynerf` for N3DV or
+`meetroom`), followed by a scene. Within each scene, modalities are grouped so
+that large subsets can be downloaded independently.
+
+For trajectory `cam00_01` and zero-based frame index `i`, use file number
+`i + 1`:
+
+```text
+images/cam00_01/cam00_01_000001.jpg
+depths/cam00_01/cam00_01_000001.jpg.geometric.png
+depth_vis/cam00_01/cam00_01_000001.jpg
+camera_params/cam00_01.json  # entry 0
+```
+
+## RGB
+
+RGB frames are 8-bit three-channel JPEG files. N3DV-derived frames have size
+1352 x 1014; MeetRoom-derived frames have size 1280 x 720.
+
+## Depth
+
+Depth files are single-channel 16-bit PNG images. A stored value `d` represents
+`d / 1000` meters. Value zero is the invalid mask and indicates negligible
+accumulated opacity. These values are rendered pseudo-labels derived from the
+reconstructed dynamic scene, not measurements from a depth sensor.
+
+```python
+from PIL import Image
+import numpy as np
+
+depth_mm = np.asarray(Image.open(depth_path), dtype=np.uint16)
+valid = depth_mm > 0
+depth_m = depth_mm.astype(np.float32) / 1000.0
+```
+
+`depth_vis` is only a colorized preview and must not be used as numeric depth.
+
+## Camera manifest
+
+Each `<trajectory>.json` is a list of 300 dictionaries in temporal order. The
+main fields are:
+
+| Field | Meaning |
+|---|---|
+| `timestamp` | Scene time in seconds. |
+| `w2c` | 4 x 4 world-to-camera homogeneous transform. |
+| `R` | 3 x 3 rotation component used by the renderer. |
+| `T` | 3-vector translation component used by the renderer. |
+| `fl_x`, `fl_y` | Focal lengths in pixels. |
+| `cx`, `cy` | Principal point in pixels. |
+| `FoVx`, `FoVy` | Horizontal and vertical field of view in radians. |
+| `width`, `height` | Image dimensions in pixels. |
+| `image_name` | Source reference-camera identifier; not an output filename. |
+
+The camera matrix is analytically prescribed along the interpolation path. It
+is exact with respect to that virtual trajectory; it is not an SfM/SLAM
+estimate. Load the JSON record at list index `i` together with file number
+`i + 1`.
+
+## Trajectories
+
+`cam{a}_{b}` denotes the ordered interpolation from reference camera `a` to
+reference camera `b`. Because the pair is ordered, `cam00_01` and `cam01_00`
+are distinct. `cam00_00` is a static-viewpoint temporal sequence. Every
+trajectory contains 300 frames spanning the scene's 10-second temporal extent.
+
+## Preview videos
+
+`video/<trajectory>_rgb.mp4` and `video/<trajectory>_depth.mp4` are provided for
+quick browsing at 30 fps. They are convenience derivatives; image sequences
+and camera JSON are the authoritative data.
+