---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: action
    dtype: 
      class_label:
        names:
          '0': None
          '1': Waving
          '2': Pointing
          '3': Clapping
          '4': Follow
          '5': Walking
          '6': Stop
          '7': Turn
          '8': Jumping
          '9': Come here
          '10': Calm
  - name: camera
    dtype: int64
  - name: subject
    dtype: int64
  - name: idx
    dtype: int64
  - name: label
    dtype: string
  - name: link
    dtype: string
  splits:
  - name: train
  - name: val
license: mit
tags:
- computer vision
- machine learning
- video understanding
- classification
- human-machine-interaction
- human-robot-interaction
- human-action-recognition
task_categories:
- video-classification
language:
- en
pretty_name: University of Technology Chemnitz - Human Robot Interaction Dataset
---

<img src="https://cdn-uploads.huggingface.co/production/uploads/6792a78d9967b7f195400ea2/JrWkycuOahvZ2AQqD9UbU.png" alt="drawing" width="200"/>

University of Technology Chemnitz, Germany<br>
Department Robotics and Human Machine Interaction<br>
Author: [Robert Schulz](mailto:robert.schulz@s2021.tu-chemnitz.de?subject=TUC-AR%20Dataset%20-%20HuggingFace)

# TUC-HRI Dataset Card

TUC-AR is an action recognition dataset, containing 10(+1) action categories for human machine interaction. This version contains video sequences, stored as images, frame by frame.

We introduce two validation types: random validation and cross-subject validation. This is the **random validation** dataset. For cross-subject validation, please use https://huggingface.co/datasets/SchulzR97/TUC-HRI-CS.
- In **random validation**, a train and a validation split are obtained by randomly splitting the sequences while maintaining an allocation rate of approximately 80\% train / 20\% validation. This ensures that each action, subject, and camera, as well as the overall number of sequences, are distributed in this ratio among the splits. Thus, we obtained 17,263 train sequences and 4,220 validation sequences.
- For **cross-subject validation**, subject _0_ and _8_ were chosen as validation subjects. All other subjects were assigned to the train split.

## Dataset Details
- RGB and depth input recorded by Intel RealSense D435 depth camera
- 12 subjects
- 11,031 sequences (train 8,893/ val 2,138)
- 3 perspectives per scene
- 10(+1) action classes
  | Action | Label     |
  |--------|-----------|
  | A000   | None      |
  | A001   | Waving    |
  | A002   | Pointing  |
  | A003   | Clapping  |
  | A004   | Follow    |
  | A005   | Walking   |
  | A006   | Stop      |
  | A007   | Turn      |
  | A008   | Jumping   |
  | A009   | Come here |
  | A010   | Calm      |

## How to Use this Dataset

1. Install the RSProduction Machine Learning package ([PyPi](https://pypi.org/project/rsp-ml/), [GitHub](https://github.com/SchulzR97/rsp-ml))
```bash
pip install rsp-ml
```
2. Use the HF datasat with `rsp.ml.dataset.TUCHRI`
```python
from rsp.ml.dataset import TUCHRI
import rsp.ml.multi_transforms as multi_transforms
import torchvision.transforms as transforms

USE_DEPTH_DATA = True

class ToNumpy:
  def __call__(self, x):
    if isinstance(x, Image.Image):
      return np.array(x)
    elif isinstance(x, torch.Tensor):
      return x.permute(1, 2, 0).numpy()  # Tensor (C, H, W) -> (H, W, C)
    else:
      raise TypeError("Input must be a PIL.Image or torch.Tensor")

transform = transforms.Compose([
  transforms.Resize((600, 600)),
  transforms.ColorJitter(brightness=0.8, contrast=0.8, saturation=0.8, hue=0.5),
  transforms.RandomRotation(180, expand=True),
  transforms.CenterCrop((375, 500)),
  #transforms.RandomCrop(input_size),
  #transforms.ToTensor(),
  ToNumpy()
])
dtd_dataset = torchvision.datasets.DTD(download=True, split='val', transform=transform)

tranforms_train = multi_transforms.Compose([
  multi_transforms.ReplaceBackground(
      backgrounds = backgrounds,
      hsv_filter=[(69, 87, 139, 255, 52, 255)],
      p = 0.8
  ),
  multi_transforms.Resize((400, 400), auto_crop=False),
  multi_transforms.Color(0.1, p = 0.2),
  multi_transforms.Brightness(0.7, 1.3),
  multi_transforms.Satturation(0.7, 1.3),
  multi_transforms.RandomHorizontalFlip(),
  multi_transforms.GaussianNoise(0.002),
  multi_transforms.Rotate(max_angle=3),
  multi_transforms.Stack()
])
transforms_val = multi_transforms.Compose([
  multi_transforms.Resize((400, 400), auto_crop=False),
  multi_transforms.Stack()
])

ds_train = TUCHRI(
  phase='train',
  load_depth_data=True,
  sequence_length=30,
  num_classes=11,
  transforms=tranforms_train
)
ds_val = TUCHRI(
  phase='val',
  load_depth_data=True,
  sequence_length=30,
  num_classes=11,
  transforms=transforms_val
)
```

## Dataset Card Contact

In case of any doubts about the dataset preprocessing and preparation, please contact [TUC RHMi](mailto:robert.schulz@s2021.tu-chemnitz.de?subject=TUC-AR%20Dataset%20-%20HuggingFace).