README.md

---
license: cc-by-4.0
task_categories:
- robotics
- image-segmentation
- image-text-to-text
---

# OSMa-Bench Dataset

[**Project Page**](https://be2rlab.github.io/OSMa-Bench/) | [**Paper**](https://huggingface.co/papers/2503.10331) | [**Code**](https://github.com/be2rlab/OSMa-Bench)

[![OSMa-Bench pipeline](https://be2rlab.github.io/OSMa-Bench/static/images/full_pipeline.png)](https://be2rlab.github.io/OSMa-Bench/)

OSMa-Bench (Open Semantic Mapping Benchmark) dataset is a fully automatically generated dataset for evaluating the robustness of open semantic mapping and segmentation systems under varying indoor lighting conditions and robot movement dynamics. This dataset is part of [OSMa-Bench](https://be2rlab.github.io/OSMa-Bench/) pipeline.

## Dataset Summary

This dataset provides simulated RGB-D and semantically annotated posed sequences for evaluation of semantic mapping and segmentation, with a particular focus on handling dynamic lighting—a critical but often overlooked factor in existing benchmarks. It also includes a collection of automatically generated question–answer pairs across multiple categories to support the evaluation of scene-graph–based reasoning, offering a task-driven measure of how well a system’s reconstructed scene captures semantic relationships between objects.

The data is built upon two base datasets:
- **ReplicaCAD**: 22 scenes with 4 lighting configurations and a velocity modifier.
- **Habitat Matterport 3D (HM3D)**: 8 scenes with 2 lighting configurations and a velocity modifier.

## Installation

We offer two versions of the dataset: one with separate files and one as a single compressed archive. Use the following command to download the separated files (may be slow):

```bash
git xet install
git clone https://huggingface.co/datasets/warmhammer/OSMa-Bench_dataset -b main
```
and this command to download the compressed version (faster one):

```bash
git xet install
git clone https://huggingface.co/datasets/warmhammer/OSMa-Bench_dataset -b compressed
unzip data.zip
```

## Data Configurations

The dataset includes the following configurations for the ReplicaCAD and HM3D scenes:

| Configuration | Description |
| :--- | :--- |
| `baseline` | Static, non-uniformly distributed light sources (ReplicaCAD only) |
| `dynamic_lighting` | Lighting conditions change along the robot's path (ReplicaCAD only) |
| `nominal_lights` | The mesh itself emits light without added light sources |
| `camera_light` | An extra directed light source is attached to the camera |
| `velocity` | Sequences recorded at doubled nominal velocity |

---

## Data Structure

The dataset provides structured data for each scene, suitable for tasks like 3D scene understanding, visual question answering, and robotics. Each scene contains the following components:

| Component                 | Description                                                                                       | Format / Example                      |
| ------------------------- | ------------------------------------------------------------------------------------------------- | --------------------------------------|
| **RGB Images**            | Standard color images captured from different camera viewpoints.                                  | `frame000000.jpg`, ...                |
| **Depth Images**          | Depth maps aligned with RGB images. Each pixel encodes depth in meters.                           | `depth000000.png`, ...                |
| **Semantic Masks**        | Pixel-wise semantic segmentation labels. Each pixel corresponds to a semantic class ID.           | `semantic000000.png`, ...             |
| **Camera Trajectories**   | Flattened 4×4 transformation matrices representing camera poses for each frame.                   | `traj.txt` (one 4×4 matrix per line)  |
| **Question-Answer Pairs** | Validated question-answer pairs related to the scene, optionally associated with specific frames. | `validated_questions.json`            |


## VQA Question Categories
The dataset includes a automatically generated answer-question pairs with the following question types:
1. **Binary General** – Yes/No questions about the presence of objects and general scene characteristics  
   *Example:* `Is there a blue sofa?`

2. **Binary Existence-Based** – Yes/No questions designed to track false positives by querying non-existent objects  
   *Example:* `Is there a piano?`

3. **Binary Logical** – Yes/No questions with logical operators such as AND/OR  
   *Example:* `Is there a chair AND a table?`

4. **Measurement** – Questions requiring numerical answers related to object counts or scene attributes  
   *Example:* `How many windows are present?`

5. **Object Attributes** – Queries about object properties, including color, shape, and material  
   *Example:* `What color is the door?`

6. **Object Relations (Functional)** – Questions about functional relationships between objects  
   *Example:* `Which object supports the table?`

7. **Object Relations (Spatial)** – Queries about spatial placement of objects within the scene  
   *Example:* `What is in front of the staircase?`

8. **Comparison** – Questions that compare object properties such as size, color, and position  
   *Example:* `Which is taller: the bookshelf or the lamp?`


## Citation

Using OSMa-Bench dataset in your research? Please cite following paper: [OSMa-Bench arxiv](https://arxiv.org/abs/2503.10331).

```bibtex
@inproceedings{popov2025osmabench,
    title     = {OSMa-Bench: Evaluating Open Semantic Mapping Under Varying Lighting Conditions},
    author    = {Popov, Maxim and Kurkova, Regina and Iumanov, Mikhail and Mahmoud, Jaafar and Kolyubin, Sergey},
    booktitle = {2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
    year      = {2025}
}
```