Back to Feed
Computer Vision / Benchmarks & Evals

Improving Landmine Detection with Better Data

Original: SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift
By Sagar Lekhak, Prasanna Reddy Pulakurthi, Lalit Joshi, Ramesh Bhatta, Emmett J. Ientilucci Published 2026-07-31

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • The authors identified and corrected seven types of errors in the original SULAND_v1 dataset, such as missing labels and incorrectly positioned bounding boxes.
  • Correcting the data significantly improved detection accuracy, with some models seeing a performance boost of nearly 20 percentage points.
  • The study highlights that high performance in controlled testing settings is not a reliable indicator of how well a model will perform in the field.
  • A new benchmark, SULAND_v2, was established using 35 different model configurations to test how robots handle new, unseen environments.

Summary & Methodology Analysis

The researchers began by auditing the existing SULAND_v1 dataset, which contained numerous flaws that hindered reliable AI training. They performed a frame by frame manual review of thousands of images, using a custom script to ensure that the labels (the digital boxes drawn around landmines) matched the actual contents of the images. By categorizing these errors, such as missing labels or mislabeled classes, they were able to create a cleaned and unified version of the data called SULAND_v2. This process preserved the original images while ensuring the labels were accurate and consistent for every sample.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main problem with existing landmine detection tools?

Existing tools often rely on low-quality datasets with errors, and they struggle to maintain accuracy when moved from a training environment to a new, unfamiliar location.

Q2. How did the researchers fix these problems?

They manually checked and corrected thousands of image labels to ensure the training data was accurate and consistent.

Q3. Why is this research important for safety?

It provides a more reliable way to test if detection robots can actually work in the real world, rather than just performing well in simplified laboratory tests.

Q4. What specifically was wrong with the original dataset labels?

The original dataset had missing labels, false identifications, incorrectly placed boxes, inconsistent rules for visible targets, and a mismatched system for labeling different types of mines.

Q5. What is a domain shift in this context?

A domain shift occurs when a robot is trained in one set of conditions but must operate in a different environment, such as a different season, terrain, or weather.

Q6. How many different AI models were tested?

The researchers evaluated 35 different model configurations across nine different groups of detector designs.

Q7. What metrics were used to judge the models?

The researchers used accuracy, precision (how many detected objects were actually mines), recall (how many total mines were found), processing speed, and the number of model parameters.

Q8. Did the study use simulated data?

The study mentions that some previous datasets used simulated data, but this research focused on refining and benchmarking with the provided SULAND dataset.

Q9. What are the primary limitations of the current study?

The study is limited by the fact that the available files do not perfectly match the original reported counts, the data has some imbalance between target and background, and collecting real mine imagery is difficult due to safety and security concerns.