---
language:
- en
tags:
- multimodal-learning
---
# Dataset Card for TST-Replica

## Dataset Description

- **Homepage:** [https://tst-vision.epfl.ch](https://tst-vision.epfl.ch)
- **Repository:** [TST official repository](https://github.com)
- **Paper:** [Arxiv](https://arxiv.org)

### Dataset Summary

This custom TST-Replica dataset is used in research work "Multimodality as Supervision: Self-Supervised Specialization to the Test Environment via Multimodality". 

- `pretrain/` is a multimodal pretraining dataset collected using Replica simulation environment. It contains RGB images, and 9 additional tokenized modalities. 

- `segmentation/train` is the associated downstream dataset used to finetune TST pretrained models on semantic segmentation tasks. 

- `segmentation/test` contains the test dataset used for evaluation/testing on semantic segmentation task. This data corresponds to samples obtained from the test-space itself.


## Dataset Structure Pretraining Data

```python
TST-Replica/
├── pretrain/
│   ├── test_spaces/
│   │   ├── crop_settings/               # Contains .tar shards
│   │   ├── det/                         # Contains .tar shards
│   │   ├── rgb/                         # Contains .tar shards
│   │   ├── tok_canny_edge@224/          # Contains .tar shards
│   │   ├── ...                          # More tokenized feature directories
│   │   └── tok_semseg@224/              # Contains .tar shards
│   └── transfer/
│       ├── crop_settings/               # Contains .tar shards
│       ├── det/                         # Contains .tar shards
│       ├── rgb/                         # Contains .tar shards
│       ├── tok_canny_edge@224/          # Contains .tar shards
│       ├── ...                          # More tokenized feature directories
│       └── tok_semseg@224/              # Contains .tar shards
├── segmentation/
│   ├── train/                          # Training data for segmentation
│   └── test/                           # Test data for segmentation
└── README.md

```

## Dataset Creation

We use [Omnidata](https://omnidata.vision), to densely
sample Replica meshes corresponding to the 5 scenes to
build our pre-training dataset. We defer the details
of the sampling procedure to Omnidata. 

### Source Data

Original dataset samples are collected from Omnidata framework. 


### Citation Information

```
@inproceedings{singh2026tst,
            title={Multimodality as Supervision: Self-Supervised Specialization to the Test Environment via Multimodality},
            author={Kunal Pratap Singh and Ali Garjani and Rishubh Singh and Muhammad Uzair Khattak and Efe Tarhan and Jason Toskov and Andrei Atanov and O{\u{g}}uzhan Fatih Kar and Amir Zamir},
            booktitle={International Conference on Learning Representations (ICLR)},
            year={2026}
        }
```