DMAT Atmospheric Turbulence Mitigation and Object Detection Dataset

These synthetic datasets are used in our paper: 

Paul Hill et al, DMAT: An End-to-End Framework for Joint Atmospheric Turbulence Mitigation and Object Detection, WACV2026

As ground truth data for atmospheric turbulence is unavailable, we generate synthetic AT distortions using the Phase-to-Space (P2S) Transform, applying them to both static and dynamic datasets designed for object detection.

Static scenes

We use the COCO2017 dataset, and to explore different use cases, we define three application subsets:

1. All: All 80 COCO categories;
2. Top10: The 10 most frequent categories—Person, Car, Chair, Book, Bottle, Cup, Dining Table, Bowl, Skis, and Handbag;
3. CarPerson: Only the Car and Person classes, relevant to surveillance tasks.

Dynamic scenes

We use the GOT-10k dataset. As the GOT-10k dataset labels only one object per frame, we generate pseudo ground truth using the YOLOv11x detector

More detail: https://arxiv.org/abs/2507.04323