SteerViT Training Data

Paper | Project Page | GitHub | Model Weights | Demo

This Hugging Face dataset hosts the referential segmentation annotations used to train SteerViT. It contains the training examples, English referential expressions, and segmentation masks.

The source images are not redistributed in this dataset due to licencing. To train SteerViT, download the required image datasets separately and point the training config to their local paths.

Dataset Contents

Each example contains:

• image_id: image identifier used to resolve the local image file.
• dataset: source dataset name.
• referential_expression: English text prompt describing the target region.
• mask: generated segmentation mask for the referred target.
• bbox: bounding box annotation from original dataset

The released SteerViT training code uses the train and validation splits.

from datasets import load_dataset

train_ds = load_dataset("JonaRuthardt/SteerViT", split="train")
val_ds = load_dataset("JonaRuthardt/SteerViT", split="validation")

print(train_ds[0].keys())

Required Local Images

The annotations reference images from:

• MS COCO 2014 train/validation images
• Visual Genome images
• Mapillary images

Only the image files are needed locally; the masks and referential expressions are provided by this Hugging Face dataset.

MS COCO

Download the 2014 train and validation image archives from the official COCO download page:

• https://cocodataset.org/#download
• http://images.cocodataset.org/zips/train2014.zip
• http://images.cocodataset.org/zips/val2014.zip

Example setup:

mkdir -p /path/to/steervit_data/coco
cd /path/to/steervit_data/coco

wget -c http://images.cocodataset.org/zips/train2014.zip
wget -c http://images.cocodataset.org/zips/val2014.zip

unzip train2014.zip
unzip val2014.zip

This should create:

/path/to/steervit_data/coco/
  train2014/
    COCO_train2014_*.jpg
  val2014/
    COCO_val2014_*.jpg

Visual Genome

Download the Visual Genome image archives from the official Visual Genome download page:

• https://homes.cs.washington.edu/~ranjay/visualgenome/api.html
• https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip
• https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip

Example setup:

mkdir -p /path/to/steervit_data/visual_genome
cd /path/to/steervit_data/visual_genome

wget -c https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip
wget -c https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip

unzip images.zip
unzip images2.zip

This should create:

/path/to/steervit_data/visual_genome/
  VG_100K/
    *.jpg
  VG_100K_2/
    *.jpg

Mapillary

Download the required Mapillary images from the Mapillary dataset page:

• https://www.mapillary.com/dataset/vistas

The Mapillary page may require logging in and accepting the dataset terms before download. After extracting the images, create a flat image directory because the SteerViT loader resolves Mapillary images as:

<mapillary_path>/<image_id>.jpg

The final directory should look like:

/path/to/steervit_data/mapillary/images/
  *.jpg

Expected File Structure

After downloading the source images, the full local image layout should look like:

/path/to/steervit_data/
  coco/
    train2014/
      COCO_train2014_000000000009.jpg
      ...
    val2014/
      COCO_val2014_000000000042.jpg
      ...

  visual_genome/
    VG_100K/
      1.jpg
      ...
    VG_100K_2/
      2411592.jpg
      ...

  mapillary/
    images/
      <mapillary_image_id>.jpg
      ...

Then point configs/data.yaml in the SteerViT repository to these roots:

dataset: "JonaRuthardt/SteerViT"
mscoco: "/path/to/steervit_data/coco"
visual_genome: "/path/to/steervit_data/visual_genome"
mapillary: "/path/to/steervit_data/mapillary/images"

Use with SteerViT Training

Clone the SteerViT repository and install it in editable mode:

git clone https://github.com/JonaRuthardt/SteerViT.git
cd SteerViT
python -m pip install -e .

After editing configs/data.yaml, launch training from the repository root:

python train.py \
  --config configs/default.yaml \
  --data_config configs/data.yaml

The training script loads the annotations automatically from this Hugging Face dataset.

Licensing and Terms

This Hugging Face dataset does not redistribute the source images. Users are responsible for downloading the underlying datasets from their official sources and complying with their licenses and terms of use:

• MS COCO: https://cocodataset.org/#terms
• Visual Genome: https://homes.cs.washington.edu/~ranjay/visualgenome/api.html
• Mapillary: https://www.mapillary.com/dataset/vistas

Citation

If you use this dataset or SteerViT in your research, please cite:

@misc{ruthardt2026steervit,
      title={Steerable Visual Representations},
      author={Jona Ruthardt and Manu Gaur and Deva Ramanan and Makarand Tapaswi and Yuki M. Asano},
      journal={arXiv:2604.02327},
      year={2026}
}