SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift
Abstract
RGB imagery offers a practical, low-cost option for Unmanned Aerial/Ground Vehicle (UAV/UGV) survey support in surface-landmine detection, but object detectors remain underexplored in this safety-critical domain. Limited cross-architecture benchmarking and insufficient out-of-distribution (OOD) analysis obscure whether detectors generalize across deployment conditions. This challenge is amplified by the scarcity of public RGB landmine datasets, making SULAND a key benchmark for PFM-1 and PMA-2 detection. However, inspection reveals missing/false annotations, localization errors, inconsistent visibility criteria, visual artifacts, temporal labeling inconsistencies, and an inverted OOD class-ID convention in SULAND. We present SULAND_v2, a refined RGB surface-landmine dataset and benchmark. Preserving original images and splits, we manually revise annotations to ensure completeness, precise localization, label validity, and class consistency. SULAND_v2 contains 33,771 images and 12,433 bounding boxes. We benchmark 35 detector configurations across nine families. Annotation refinement improves YOLOv8 in-distribution (IID) test mAP@50 by 14.6-19.6 percentage points, while fixing the OOD class-ID convention increases mean YOLOv8 OOD mAP@50 by ~25 percentage points. On SULAND_v2, YOLOv12-Small achieves the highest IID mAP@50 (0.908), while RF-DETR-Large yields the strongest OOD performance (0.799 mAP@50, 0.675 recall). Our results demonstrate that high IID accuracy does not guarantee operational readiness. SULAND_v2 provides a reliable benchmark for evaluating domain-shift robustness in RGB-based mine-action survey support.
Community
We release SULAND_v2, a refined RGB surface-landmine dataset and an IID/OOD object-detection benchmark for UAV/UGV-based mine-action survey support. Public RGB landmine datasets are scarce, so each one disproportionately shapes which methods get trusted. Auditing SULAND frame-by-frame turned up missing and false annotations, mislocalized boxes, inconsistent partial-visibility criteria, and, critically, an inverted class-ID convention in the OOD split. We manually re-annotated the whole dataset in Label Studio (33,771 images, 12,433 boxes), keeping the original imagery and splits intact so v1 and v2 stay directly comparable. What the refinement is worth, quantitatively: YOLOv8 IID test mAP@50 improves by 14.6 to 19.6 pp, and fixing the class-ID convention alone adds roughly 25 pp to mean OOD mAP@50. Cross-version training and evaluation (v1→v2, v2→v1, and so on) separates training-label quality from evaluation-ground-truth consistency. On the refined benchmark we train 35 configurations across 9 detector families (one-stage, two-stage, transformer, open-vocabulary) under one protocol. YOLOv12-S leads IID (0.908 mAP@50); RF-DETR-L leads OOD (0.799 mAP@50, 0.675 recall). The ranking reshuffles entirely under domain shift, and model scale does not predict robustness. High IID accuracy is not evidence of operational readiness.
Dataset: https://huggingface.co/datasets/SagarLekhak/SULAND_v2_RGB_Surface_Landmine_Dataset
Code: https://github.com/PrasannaPulakurthi/SULAND_v2
Get this paper in your agent:
hf papers read 2607.28996 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 1
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper