Introduction
In aerial surveillance and defense, detecting "small and dim" targets—such as distant missiles or aircraft plumes—remains a formidable challenge. These targets typically occupy only a few pixels, exhibit low contrast, and are often obscured by heavy background clutter. While several Infrared Search and Track (IRST) datasets exist, they frequently rely on synthetic overlays or uncooled Long-Wave Infrared (LWIR) sensors. The MWIRSTD dataset addresses these limitations by providing the first publicly available collection of real-world Mid-Wave Infrared (MWIR) video sequences captured with high-sensitivity cooled imagers.
The MWIRSTD Dataset
The primary contribution of this work is the MWIRSTD dataset, which consists of 14 video sequences totaling 1,053 images. Unlike previous benchmarks that use "target-over-background" digital overlays, MWIRSTD features authentic thermal signatures captured in outdoor environments.
The dataset includes three target classes: cracker rockets (simulating fast-moving thermal targets), debris, and moving ground vehicles. Scenarios were staged at distances ranging from 300m to 950m against complex backgrounds like mountains, vegetation, and roads. The use of a cooled Stirling-cycle InSb (Indium Antimonide) imager ensures high fidelity, capturing the 3–5 $\mu$m spectral band where high-temperature military targets are most prominent.
Methodology and Benchmarking
The researchers established a benchmarking pipeline to compare two generations of computer vision:
- Traditional Morphological Algorithms: Six methods based on local contrast, including Top-hat transform, Maxmean, and DGRAD.
- Deep Learning Architectures: Two state-of-the-art segmentation models, U-Net and DNA-Net, specifically adapted for small target detection.
Key Results and Insights
The experimental results reveal a significant "reality gap." Traditional algorithms that often perform well on synthetic data struggled immensely with the real-world clutter in MWIRSTD.
Deep learning models demonstrated clear dominance:
- U-Net achieved the highest performance with an Intersection over Union (IoU) of 0.79 and a Probability of Detection (PD) of 0.8855.
- DNA-Net, despite being a specialized small-target network, followed with an IoU of 0.64.
- Traditional Methods failed to generalize, with IoU scores near 0.003, often producing fragmented detections or failing to distinguish targets from sensor noise.
The study emphasizes that MWIR is the "sweet spot" for defense applications because it captures high-energy plumes more effectively than LWIR. Furthermore, the sequential nature of the data suggests that future research should leverage spatio-temporal regularity through "tracking-before-detection" (TBD) algorithms.
Conclusion
MWIRSTD provides a critical benchmark for the IRST community, proving that deep learning is essential for navigating the complexities of real-world MWIR environments. While the dataset is currently focused on frame-by-frame detection, its video-based structure opens doors for advanced temporal filtering and multi-frame tracking research.
The dataset and associated code are available at https://github.com/avinres/MWIRSTD.
Summary updated