---
pretty_name: IGVC Segmentation Dataset
tags:
  - image
license: cc-by-4.0
---

# IGCV Segmentation Dataset

Dataset for training a semantic image segmentation model for the [Intelligent Ground Vehicle Competition](http://www.igvc.org/).

## Composition

Each instance consists of an reference image from the point of view of the robot and the corresponding obstacle (e.g. construction drums, buckets) and lane segmentation masks.

**Train**

256 frames rendered in 4 different lighting environments using Blender = 1024 images 

**Test**

10 frames captured from the [SCR 2023 IGVC run](https://www.youtube.com/watch?v=7tZsk3T3STA) (manually segmented) + 13 frames rendered in 4 lighting environments = 62 images

## Usage

For usage with PyTorch it is recommended to wrap the dataset into a [`Dataset`](https://docs.pytorch.org/docs/stable/data.html#torch.utils.data.Dataset) adapter class and generate a training/validation split:

```python
from torch.utils.data import Dataset
from datasets import load_dataset, Dataset as HFDataset
import numpy as np

class Split:
    TRAIN = "train"
    VALID = "valid"
    TEST = "test"

class SegmentationDataset(Dataset):
    def __init__(self, path="Nico0302/IGVC-Segmentation", split=Split.TRAIN, transform=None, mask_name="obstacle_mask", valid_size=0.125):
        self.path = path
        self.split = split
        self.transform = transform
        self.mask_name = mask_name
        self.valid_size = valid_size

        self.data = self._read_split()

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        item = self.data[idx]

        sample = dict(image=np.array(item["image"]), mask=np.array(item[self.mask_name]))
        if self.transform is not None:
            sample = self.transform(**sample)

        return {
            "image": np.transpose(sample["image"], (2, 0, 1)), # HWC to CHW (3, H, W)
            "mask": np.expand_dims(sample["mask"].astype(np.float32) / 255.0, 0), # HW to CHW (1, H, W)
        }

    def _read_split(self):
        dataset = load_dataset(self.path, split="test" if self.split == Split.TEST else "train")
        assert isinstance(dataset, HFDataset), "Dataset must be a Hugging Face Dataset"

        if (self.split == Split.TEST):
            return dataset
        
        splits = dataset.train_test_split(test_size=self.valid_size, seed=42)
        if self.split == Split.VALID:
            return splits["test"]
        return splits["train"]
```

Using this adapter, the dataset can simple be passed to the [`DataLoader`](https://docs.pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader):

```python
train_dataset = SegmentationDataset(split=Split.TRAIN)
valid_dataset = SegmentationDataset(split=Split.VALID)
test_dataset = SegmentationDataset(split=Split.TEST)

train_dataloader = DataLoader(train_dataset)
valid_dataloader = DataLoader(valid_dataset)
test_dataloader = DataLoader(test_dataset)
```

## Acknowledgements  

Thank you for [Sooner Competitive Robotics](https://ou.edu/scr/) for allowing me to use frames from their IGVC 2023 run video as part of the test set.

## Citation

If you are using this dataset, please cite

```bibtex
@misc{gres2025IGVC,
    author       = { Nicolas Gres },
    title        = { IGCV Segmentation Dataset },
    year         = 2025,
    url          = { https://huggingface.co/datasets/Nico0302/IGVC-Segmentation },
    publisher    = { Hugging Face }
}
```