robotflowlabs/anima-thermal-synthetic

---
license: cc-by-4.0
task_categories:
- image-to-image
- depth-estimation
- object-detection
- image-segmentation
tags:
- thermal-imaging
- synthetic-data
- robotics
- adverse-weather
- night-vision
- fog-detection
- sensor-fusion
- slam
- defense
- autonomous-navigation
- data-augmentation
- computer-vision
- anima
- robot-flow-labs
pretty_name: "ANIMA Thermal Synthetic — RGB-to-Thermal Dataset for Adverse-Condition Robotics"
size_categories:
- 10K<n<100K
language:
- en
---

<div align="center">

# ANIMA Thermal Synthetic Dataset

### See Through Fog, Rain, Smoke, and Total Darkness

**27GB | 30,781 paired RGB-to-Thermal images | Production-ready**

*Built by [Robot Flow Labs](https://robotflowlabs.com) / AIFLOW LABS LIMITED*

Part of the **ANIMA Perception Suite** — 96 AI modules for robotics

</div>

---

## The Problem

Thermal cameras see what RGB cameras cannot: through fog, rain, smoke, and complete darkness. But thermal training data is scarce and expensive:

- Radiometric thermal cameras cost **$5,000+**
- Existing public thermal datasets have **fewer than 10K images**
- No large-scale multi-domain synthetic thermal dataset existed — until now

## The Solution

We trained a VLM-guided thermal synthesis model (SOL/TherA) on real **FLIR Boson+ 640x512** thermal-depth pairs, then generated **30,781 synthetic thermal images** from COCO at scale.

The result: the largest open synthetic thermal dataset for robotics and defense applications.

---

## Sample Previews

### Real Thermal Reference (VIVID++ — FLIR Boson+ 640x512)

These samples show the real thermal ground truth we trained our generator against:

| | RGB Input | Our Synthetic Thermal | Real Thermal (FLIR Boson+) |
|:-:|:---------:|:---------------------:|:--------------------------:|
| 1 |  |  |  |
| 2 |  |  |  |
| 3 |  |  |  |

### COCO Synthetic Thermal (large-scale generation)

| | RGB Input | Synthetic Thermal |
|:-:|:---------:|:-----------------:|
| 1 |  |  |
| 2 |  |  |
| 3 |  |  |

---

## Dataset at a Glance

| Property | Value |
|----------|-------|
| **Total images** | 30,781 paired RGB + Thermal |
| **Total size** | 27 GB |
| **Train** | 30,338 COCO + 399 VIVID++ |
| **Validation** | 44 VIVID++ (with real thermal ground truth) |
| **Format** | PNG (lossless) |
| **License** | CC-BY-4.0 |

## Data Sources

| Domain | Source | Count | Purpose |
|--------|--------|-------|---------|
| General Objects | COCO 2017 | 30,338 | Diverse objects, scenes, lighting |
| Real Thermal | VIVID++ (FLIR Boson+) | 443 | Ground truth calibration + validation |

---

## Dataset Structure

```
anima-thermal-synthetic/
├── train/
│ ├── coco/
│ │ ├── rgb/ # 30,338 source RGB images
│ │ └── thermal_synth/ # 30,338 generated thermal images
│ └── vivid/
│ ├── rgb/ # 399 source RGB
│ ├── thermal_synth/ # 399 generated thermal
│ └── thermal_real/ # 399 real FLIR Boson+ thermal (ground truth)
└── val/
├── rgb/ # 44 validation RGB
├── thermal_synth/ # 44 generated thermal
└── thermal_real/ # 44 real thermal (ground truth)
```

---

## Quick Start

```python
from huggingface_hub import snapshot_download

# Download full dataset
snapshot_download(
"robotflowlabs/anima-thermal-synthetic",
repo_type="dataset",
local_dir="./anima-thermal"
)
```

```python
from PIL import Image
from pathlib import Path

# Load a paired sample
data = Path("./anima-thermal/train/coco")
rgb = Image.open(data / "rgb" / "000000000009.png")
thermal = Image.open(data / "thermal_synth" / "000000000009.png")
```

---

## Use Cases

**Robotics** — Train depth estimation, SLAM, and navigation models for night/fog/rain conditions without expensive thermal cameras

**Defense & Security** — Thermal detection in zero-visibility: surveillance, perimeter security, UAV operations

**Industrial** — Thermal anomaly detection for manufacturing, predictive maintenance, quality inspection

**Research** — RGB-to-thermal domain adaptation, multi-modal learning, data augmentation

---

## Part of ANIMA

This dataset powers multiple modules in the [ANIMA robotics perception stack](https://github.com/RobotFlow-Labs):

| Module | How it uses this dataset |
|--------|------------------------|
| **NOTT** | Thermal SLAM — depth estimation from thermal |
| **SOL** | Generator validation and quality benchmarking |
| **FREYA** | LiDAR-Camera-IMU fusion with thermal augmentation |
| **LOKI** | Object detection robustness in night/fog |

---

## Citation

```bibtex
@dataset{robotflowlabs2026thermal,
title={ANIMA Thermal Synthetic: RGB-to-Thermal Dataset for Adverse-Condition Robotics},
author={Robot Flow Labs},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/datasets/robotflowlabs/anima-thermal-synthetic}
}
```

## License

**CC-BY-4.0** — Free for research and commercial use with attribution.

Source acknowledgments: VIVID++ (Lee et al., IEEE RA-L 2022), COCO (Microsoft, CC BY 4.0)

---

<div align="center">

**Robot Flow Labs** — *96 modules. Every sensor. Any robot. Anywhere.*

[GitHub](https://github.com/RobotFlow-Labs) · [HuggingFace](https://huggingface.co/robotflowlabs)

</div>