---
pretty_name: Label Factory Multi-Task RGB-D Collection
license: mit
tags:
- image
- video
- 3d
- rgb-d
configs:
- config_name: depth_estimation
  data_files:
  - split: train
    path: "*/data/depth_estimation/train-*.parquet"
  - split: validation
    path: "*/data/depth_estimation/validation-*.parquet"
  - split: test
    path: "*/data/depth_estimation/test-*.parquet"
- config_name: instance_segmentation
  data_files:
  - split: train
    path: "*/data/instance_segmentation/train-*.parquet"
  - split: validation
    path: "*/data/instance_segmentation/validation-*.parquet"
  - split: test
    path: "*/data/instance_segmentation/test-*.parquet"
- config_name: object_pose_estimation
  data_files:
  - split: train
    path: "*/data/object_pose_estimation/train-*.parquet"
  - split: validation
    path: "*/data/object_pose_estimation/validation-*.parquet"
  - split: test
    path: "*/data/object_pose_estimation/test-*.parquet"
- config_name: object_pose_models
  data_files:
  - split: train
    path: "*/data/object_pose_estimation/models.parquet"
---

# KarolyArmin/Hand_tools

Growing multi-task RGB-D collection generated with the Label Factory workflow.
Every capture is stored in its own zero-padded directory (`0000`, `0001`,
`0002`, ...). Uploading a new capture appends it to the collection without
replacing earlier directories. Task configurations aggregate compatible
Parquet shards from every numbered directory.

## Training configurations

- `depth_estimation`: RGB input, metric depth target, intrinsics and depth units
- `instance_segmentation`: RGB input, instance-mask target, boxes and annotations
- `object_pose_estimation`: RGB-D input, masks, camera transforms and object poses
- `object_pose_models`: one STL payload per object type and numbered dataset

Every configuration provides deterministic `train`, `validation`, and `test`
splits. Adding another numbered capture appends rows through the wildcard paths;
it does not replace earlier datasets.

## Loading all or several numbered datasets

The configuration automatically combines every uploaded directory (`0000`,
`0001`, `0002`, ...). Filter `dataset_id` to use selected captures:

```python
from datasets import load_dataset

poses = load_dataset("KarolyArmin/Hand_tools", "object_pose_estimation")
selected_ids = set(["0000", "0001", "0004"])
for split_name in poses:
    poses[split_name] = poses[split_name].filter(
        lambda row: row["dataset_id"] in selected_ids
    )
```

## Depth-estimation example

```python
from datasets import load_dataset
import numpy as np

depth_dataset = load_dataset("KarolyArmin/Hand_tools", "depth_estimation")
sample = depth_dataset["train"][0]
rgb = sample["image"].convert("RGB")
depth_native = np.asarray(sample["depth"])
depth_mm = depth_native.astype(np.float32) * sample["depth_scale"]
depth_m = depth_mm / 1000.0
intrinsics_3x3 = np.asarray(sample["camera_intrinsics"]).reshape(3, 3)
```

## Instance-segmentation example

```python
from datasets import load_dataset
import json
import numpy as np

instances = load_dataset("KarolyArmin/Hand_tools", "instance_segmentation")
sample = instances["validation"][0]
rgb = sample["image"].convert("RGB")
instance_mask = np.asarray(sample["instance_mask"])
annotations = json.loads(sample["annotations_json"])
boxes_xywh = np.asarray(sample["boxes_xywh"])
instance_names = sample["instance_names"]
```

## Object-pose-estimation example

```python
from datasets import load_dataset
import numpy as np

poses = load_dataset("KarolyArmin/Hand_tools", "object_pose_estimation")
sample = poses["test"][0]
rgb = sample["image"].convert("RGB")
depth = np.asarray(sample["depth"])
instance_mask = np.asarray(sample["instance_mask"])
world_from_camera = np.asarray(sample["world_from_camera"]).reshape(4, 4)
camera_from_world = np.asarray(sample["camera_from_world"]).reshape(4, 4)
world_from_objects = np.asarray(sample["world_from_objects"]).reshape(-1, 4, 4)
camera_from_objects = np.asarray(sample["camera_from_objects"]).reshape(-1, 4, 4)
object_names = sample["object_names"]

# Resolve the STL models referenced by this pose sample.
models = load_dataset("KarolyArmin/Hand_tools", "object_pose_models")["train"]
required_keys = set(sample["object_model_keys"])
models = models.filter(lambda row: row["object_model_key"] in required_keys)
first_model = models[0]
with open(first_model["stl_file_name"], "wb") as file:
    file.write(first_model["stl"])
```

## Collection contents

Each numbered directory contributes:

- RGB images and aligned 16-bit metric depth maps
- camera intrinsics and metric world-coordinate camera poses
- instance masks, boxes, RLE annotations and visualization overlays
- per-frame and consolidated object poses
- metric scene reconstruction and object STL models
- task-specific train, validation and test Parquet shards

## Latest generated contribution

- Directory: `0002/`
- RGB-D frames: 147
- Annotated object instances: 4
- Selected upload files: 922

Depth images are stored in millimetres. Pose conventions and coordinate-system
metadata are recorded in `0002/sam6d_scene/annotations.json` and
`0002/reconstruction/metric_alignment/camera_poses_metric.json`.

## Limitations

Annotations are automatically generated and must be independently reviewed for
the intended downstream use. Reconstruction, depth-alignment, segmentation and
pose-estimation errors may remain.