SagarLekhak/SULAND_v2_RGB_Surface_Landmine_Dataset

SULAND_v2: Refined RGB Surface Landmine Detection Dataset

SULAND_v2 is a refined annotation release based on the original SULAND (SUrface LANDmine) RGB dataset for surface-landmine detection.

The original images, video folders, frame names, and sequence-level training, validation, and test splits are retained. Only the annotations have been revised. To support reproducible comparisons and reduce temporal leakage, users are recommended to preserve the provided sequence-level splits.

The motivation for SULAND_v2, the annotation issues identified in the earlier release, the refinement procedure, and the associated benchmark experiments are described in our paper:

SULAND_v2 paper: SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift

In our paper, the original SULAND release is referred to as SULAND_v1.

Classes

Annotations use the normalized YOLO bounding-box format:

<class_id> <x_center> <y_center> <width> <height>

Repository files

Recommended use

For experiments using the refined SULAND_v2 annotations, use:

SULAND_images.zip
Annotation_files_for_SULAND_v2.zip

For experiments using the original SULAND_v1 annotations, use:

SULAND_images.zip
Annotations_files_for_SULAND_v1.zip

For SULAND_v1 experiments requiring consistent IID and OOD class IDs, use:

SULAND_images.zip
Annotation_files_for_SULAND_v1_OOD_ID_corrected.zip

Only one annotation archive should be used at a time.

Download dataset

Install the Hugging Face Hub package:

pip install huggingface_hub

Download the SULAND_v2 annotations and image archive:

from huggingface_hub import hf_hub_download

repo_id = "SagarLekhak/SULAND_v2_RGB_Surface_Landmine_Dataset"

hf_hub_download(
    repo_id=repo_id,
    filename="Annotation_files_for_SULAND_v2.zip",
    repo_type="dataset",
    local_dir="./datasets/",
)

hf_hub_download(
    repo_id=repo_id,
    filename="SULAND_images.zip",
    repo_type="dataset",
    local_dir="./datasets/",
)

Prepare the SULAND_v2 dataset

Extract the annotation and image archives:

unzip ./datasets/Annotation_files_for_SULAND_v2.zip -d ./datasets
unzip ./datasets/SULAND_images.zip -d ./datasets

Merge the image folders into the extracted SULAND_v2 directory while preserving the relative folder structure:

rsync -a ./datasets/SULAND_images/ ./datasets/SULAND_v2/

The resulting dataset contains the original images together with the refined SULAND_v2 annotations.

Dataset structure

datasets/SULAND_v2/
├── data-iid/                         # In-distribution data collected in Italy
│   ├── ITA.yaml
│   ├── train/
│   │   ├── images/
│   │   └── labels/
│   ├── val/
│   │   ├── images/
│   │   └── labels/
│   ├── test/
│   │   ├── images/
│   │   └── labels/
│   └── annotations/                  # Generated COCO JSON annotations
│       ├── instances_train.json
│       ├── instances_val.json
│       └── instances_test.json
└── data-ood/                         # Out-of-distribution data collected in the USA
    ├── USA.yaml
    ├── val/
    │   ├── images/
    │   └── labels/
    └── annotations/                  # Generated COCO JSON annotations
        └── instances_val.json

Images and labels retain corresponding relative folder paths and filename stems. Users are recommended to retain the original sequence-level splits rather than randomly mixing frames across sequences.

Codebase & Benchmarks

The codebase and training scripts to reproduce our cross-version evaluations and benchmark experiments are available on GitHub:

👉 https://github.com/PrasannaPulakurthi/SULAND_v2

We evaluate the effect of dataset refinement across 35 detector configurations spanning 9 detector families:

• —YOLO Series: YOLOv8 (N/S/M/L/X), YOLO11 (N/S/M/L/X), YOLOv12 (N/S/M/L/X), and YOLO26 (N/S/M/L/X)
• —Open-Vocabulary YOLO: YOLO-Worldv2 (S/M/L/X)
• —Two-Stage CNN: Faster R-CNN (ResNet-50 / ResNet-101)
• —Real-Time Transformers: RT-DETR (L/X), D-FINE (N/S/M/L/X), and RF-DETR (Base/Large)

Citation

When using this dataset, please cite SULAND_v2 paper:

@misc{lekhak2026sulandv2refinedrgb,
      title={SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift}, 
      author={Sagar Lekhak and Prasanna Reddy Pulakurthi and Lalit Joshi and Ramesh Bhatta and Emmett J. Ientilucci},
      year={2026},
      eprint={2607.28996},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.28996}, 
}

Original SULAND attribution

The original SULAND dataset was created by Emanuele Vivoli, Marco Bertini, and Lorenzo Capineri. We acknowledge their work as the source of the images and original annotations used as the starting point for SULAND_v2.

SULAND-Dataset: https://github.com/miccunifi/SULAND-Dataset

Disclaimer

This dataset is intended for academic research. Performance on this dataset should not be interpreted as evidence of operational readiness for real-world humanitarian demining.