File size: 2,789 Bytes
9bc2a0c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 | ---
license: mit
tags:
- robotics
- pointcloud
- droid
- 3d-object-detection
---
# DROID 3D Bounding-Box Pipeline — Visualization Data
Per-frame fused 3-camera pointclouds backing the web viewer at
<https://silicon23.github.io/droid_pipeline_visualization/>.
Each episode comes from the DROID v1.0.1 dataset and completed the full
detection pipeline (FoundationStereo depth -> optimized extrinsics -> SAM3
masks -> SAM3D box -> FoundationPose video tracking). Point positions are in
the **Franka base (world) frame**, metres, Z up.
## Layout
```
clouds/{episode_id}/cloud_{preset}.bin concatenated per-frame gzip blocks
clouds/{episode_id}/index_{preset}.json byte offsets + per-frame metadata
```
Presets trade file size against density:
| preset | pixel stride | voxel | typical |
|---|---|---|---|
| `low` | 4 | 15 mm | ~230 KB/frame |
| `medium` | 4 | 10 mm | ~380 KB/frame |
| `high` | 2 | 6 mm | ~1.25 MB/frame |
## Block format
Blocks are listed in `index_{preset}.json` as `{t, o, c, n}` — source frame
index, byte offset, compressed length, point count. Fetch one with an HTTP
`Range` request and gunzip it. The inflated block is:
| offset | type | meaning |
|---|---|---|
| 0 | `float32[3]` | `lo` — quantization lower corner (world, m) |
| 12 | `float32[3]` | `hi` — quantization upper corner |
| 24 | `int16[n*3]` | XYZ, linearly mapped `lo..hi` -> `-32768..32767` |
| 24 + 6n | `uint8[n*3]` | RGB |
Dequantize with `xyz = lo + (q + 32768) / 65535 * (hi - lo)`.
`index_{preset}.json` also carries `frames[]` with the per-frame
FoundationPose `status`, `consensus_size`, and the 8-corner `bbox_world`.
## Reading a frame in Python
```python
import gzip, json, numpy as np, requests
BASE = "https://huggingface.co/datasets/Silicon23/droid_pipeline_visualization/resolve/main"
eid, preset, frame = "shard01010_ep008", "medium", 40
ix = requests.get(f"{BASE}/clouds/{eid}/index_{preset}.json").json()
b = ix["blocks"][frame]
raw = requests.get(f"{BASE}/clouds/{eid}/{ix['bin']}",
headers={"Range": f"bytes={b['o']}-{b['o']+b['c']-1}"}).content
buf = gzip.decompress(raw)
lo = np.frombuffer(buf, np.float32, 3, 0)
hi = np.frombuffer(buf, np.float32, 3, 12)
n = b["n"]
q = np.frombuffer(buf, np.int16, n * 3, 24).reshape(n, 3).astype(np.float32)
rgb = np.frombuffer(buf, np.uint8, n * 3, 24 + n * 6).reshape(n, 3)
xyz = lo + (q + 32768) / 65535 * (hi - lo)
```
## Notes
- Depth is cropped at 2.5 m to keep the robot workspace and drop far-wall clutter.
- Wrist-camera pose is per-frame from `trajectory.h5` (Euler XYZ, camera-to-world
directly). Episodes flagged `wrist_sensor_flipped` use the right lens composed
with a 180 deg rotation about Z.
- Only the **left** eye of each side-by-side stereo recording is used.
|