DSIAC MWIR Dataset: ATR Benchmark

• DSIAC MWIR dataset is a defense-oriented benchmark featuring mid-wave infrared and visible imagery captured at multiple ranges and conditions.
• The dataset provides detailed acquisition parameters, including frame counts, resolutions, and range increments, supporting robust ATR evaluations and domain transfer studies.
• Advanced preprocessing techniques, such as CycleGAN-based normalization and targeted augmentation, enable effective transductive transfer learning between MWIR and VIS modalities.

Searching arXiv for the cited DSIAC-related papers and closely related work to ground the article. The DSIAC MWIR dataset is the mid-wave infrared component of the DSIAC Automatic Target Recognition dataset, also described as the DSIAC ATR Algorithm Development Image Database and, in another characterization, the US Army NVESD ATR Algorithm Development Image Database. Across the cited studies, it is used as a defense- and surveillance-oriented benchmark for automatic target recognition, detection, and localization under substantial variation in range, aspect, modality, and operational conditions. The core modalities are mid-wave infrared (MWIR) and visible (VIS); the dataset is described as covering targets recorded from 1 km to 5 km, with multiple poses, view angles, and day/night conditions, and as being publicly accessible through DSIAC (Sami et al., 2023, Safdar et al., 15 Jul 2025).

1. Provenance and benchmark identity

The dataset is reported as having been collected by the US Army Night Vision and Electronic Sensors Directorate (NVESD), and the cited studies provide DSIAC access links for the ATR Algorithm Development Image Database. One study frames it explicitly as a two-domain ATR resource with MWIR and VIS imagery for transfer learning, while another frames it as a thermal-IR-centric benchmark for automatic target detection, localization, and recognition in the defense/surveillance domain (Sami et al., 2023, Safdar et al., 15 Jul 2025).

In the YOLOatr characterization, the MWIR stream is identified as thermal IR in the $3\text{–}5\,\mu\text{m}$ band, and the paired visible stream is also present. The same source describes two camera streams, labeled “cegr” for MWIR and “i1co” for visible, with approximate data sizes of $207$ GB and $106$ GB, respectively. The files are reported to be in ARF format, with ground truth in ATG format, and ARF is stated to be viewable via an ImageJ plugin (Safdar et al., 15 Jul 2025).

A common benchmark-adjacency misconception is confusion between DSIAC and SENSIAC. A separate long-range MWIR small-target detection paper explicitly states that it uses SENSIAC rather than DSIAC, and that no DSIAC sequences are used or referenced there (Kwan et al., 2020). This distinction matters because exact performance expectations are benchmark-specific.

2. Modalities, acquisition structure, and scene geometry

In the transductive ATR study, the dataset is used in two modalities: MWIR and VIS. The transfer-learning setup treats MWIR as the labeled source domain and VIS as the unlabeled target domain. That study reports $189$ MWIR video sequences and $97$ VIS video sequences, with each sequence containing $1{,}800$ frames and a typical frame resolution of $640\times 480$ pixels. It also gives approximate frame counts of $340{,}200$ for MWIR and $174{,}600$ for VIS (Sami et al., 2023).

The same study states that vehicles are recorded at distances from $1$ km to $207$0 km in $207$1 km increments, and that the dataset includes multiple poses and view angles. Performance is analyzed across ranges, with degradation beyond $207$2 km attributed to lower chip quality (Sami et al., 2023).

The YOLOatr characterization adds acquisition details that are not specified in the transfer-learning paper. It reports $207$3-minute videos at $207$4 fps, corresponding to $207$5 frames per video; $207$6 aspect angles per vehicle type, obtained by driving in a circle of $207$7 m radius at $207$8 mph; range collection from $207$9 m to $106$0 m in $106$1 m increments; and both day and night acquisitions in a desert environment (Safdar et al., 15 Jul 2025).

The content is described as containing small tactical targets at long standoff distances. At longer ranges, especially around $106$2 km, targets are reported to occupy only a few pixels on target, with reduced structural detail due to sensor resolution and distance. This suggests that range-induced scale collapse is a defining property of the benchmark, not merely a nuisance variable (Safdar et al., 15 Jul 2025).

3. Target classes and annotation conventions

For ATR in the MWIR-to-VIS transfer setting, ten vehicle categories are enumerated: 2S3, BTR70, BRDM2, BMP2, MT-LB, T72, ZSU23, D20, Pickup, and Sport vehicle. The first seven are identified as military vehicles, D20 as an artillery piece, and Pickup and Sport vehicle as civilian classes (Sami et al., 2023).

The YOLOatr paper describes the dataset somewhat differently. Its Table 1 specifies “10 vehicles (2 civilian and 8 tactical vehicles)” plus “2 human” classes, while other text mentions “13 different tactical and civilian vehicles.” The authors do not enumerate all classes in that paper, but figures and tables reference T-72 tank, BRDM2, BTR70, pickup truck, and SUV. The same paper does not reconcile the discrepancy between “10 vehicles” and “13 different tactical and civilian vehicles” (Safdar et al., 15 Jul 2025).

Annotation conventions also differ by task formulation. In the transductive transfer study, MWIR images are treated as labeled with class identities, enabling a well-trained source ATR classifier, while VIS images are treated as unlabeled target data during initial target-classifier construction. That study does not enumerate official bounding box annotations; instead, it states that vehicles are “detected and cropped” using information from Meta-UDA, implying the use of detection outputs for chip generation without detailing annotation quality or detector architecture (Sami et al., 2023).

By contrast, the detection-oriented YOLOatr study states that ground truth annotations are in ATG format. It does not explicitly specify the annotation schema, but notes that the object-detection setup and YOLO-based training imply bounding box coordinates with class labels per instance. Annotation quality beyond being “annotated for ground truth” is not discussed (Safdar et al., 15 Jul 2025).

4. Preprocessing and experimental protocols

In the transductive ATR workflow, vehicles are first detected in raw frames and cropped into target chips using information from Meta-UDA. All cropped targets from different ranges are then projected to a canonical distance of $106$3 km via bicubic interpolation, with a final target-chip size of $106$4. The stated rationale is to normalize apparent target scale across ranges and reduce variability due to distance; no physical camera model or explicit scaling equation is provided (Sami et al., 2023).

That study randomly partitions the dataset into train, validation, and test sets in a $106$5 ratio for both domains. It reports no explicit class balancing or augmentation strategy beyond the $106$6 km projection and CycleGAN translation (Sami et al., 2023).

The YOLOatr study adopts a different preprocessing regime. It performs video-to-frame conversion, resizes training images to $106$7, and applies a custom augmentation profile with the following reported values: $106$8, $106$9, $189$0, $189$1, translate $189$2, scale $189$3, shear $189$4, perspective $189$5, $189$6, $189$7, mosaic $189$8, mixup $189$9, and copy_paste $97$0. Albumentations-based blurs and noise are stated to be turned off (Safdar et al., 15 Jul 2025).

Its experimental subset uses $97$1 images for each of four selected vehicle types—T-72, BRDM2/BTR70, pickup, and SUV—with a $97$2 train, $97$3 validation, and $97$4 test split per vehicle type. It also formalizes two range protocols: correlated testing $97$5, in which training and testing images originate from the same ranges, and decorrelated testing $97$6, in which training is performed on DS1 $97$7 and testing on DS2 $97$8, both including day and night data (Safdar et al., 15 Jul 2025).

5. Use in ATR and transductive transfer learning

The DSIAC MWIR dataset is used in one study as the labeled source domain for unpaired transductive transfer learning from MWIR to VIS. The framework employs two generators, $97$9 and $1{,}800$0, two discriminators $1{,}800$1 and $1{,}800$2, and source and target ResNet-18 classifiers for $1{,}800$3-class ATR. The MWIR classifier is trained to high accuracy and held fixed; its weights are also used to initialize the VIS classifier (Sami et al., 2023).

The losses optimized in that framework are adversarial, cycle-consistency, identity, and categorical cross-entropy. The cycle-consistency and identity terms are reported as

$1{,}800$4

and

$1{,}800$5

with $1{,}800$6 and $1{,}800$7. The total objective includes a categorical cross-entropy term weighted by $1{,}800$8 for the first $1{,}800$9 epochs and $640\times 480$0 thereafter (Sami et al., 2023).

Training details reported for this DSIAC-based setup include PyTorch implementation, an NVIDIA RTX-8000 GPU, batch size $640\times 480$1, Adam for CycleGAN with $640\times 480$2, $640\times 480$3, learning rate $640\times 480$4 for the first $640\times 480$5 epochs and $640\times 480$6 for the next $640\times 480$7, discriminator updates $640\times 480$8 less frequent than generator updates, and initialization of generator weights from a “summer-to-winter” Yosemite CycleGAN model. The classifiers use Adam with $640\times 480$9, $340{,}200$0, learning rate $340{,}200$1, and $340{,}200$2 epochs, with an additional $340{,}200$3 epochs for semi-supervised fine-tuning when $340{,}200$4, $340{,}200$5, or $340{,}200$6 labeled VIS data are introduced (Sami et al., 2023).

This use of the dataset emphasizes not only classification performance but also cross-modal adaptation. A plausible implication is that the paired presence of MWIR and VIS streams makes DSIAC particularly suitable for studying domain shift under controlled range and pose variation.

6. Reported performance, limitations, and benchmark interpretation

On the DSIAC-based transductive ATR benchmark, the source-domain MWIR classifier achieves an overall test accuracy of $340{,}200$7. Range-wise accuracies are reported as $340{,}200$8 at $340{,}200$9 m, $174{,}600$0 at $174{,}600$1 m, $174{,}600$2 at $174{,}600$3 m, $174{,}600$4 at $174{,}600$5 m, $174{,}600$6 at $174{,}600$7 m, $174{,}600$8 at $174{,}600$9 m, $1$0 at $1$1 m, $1$2 at $1$3 m, and $1$4 at $1$5 m. The corresponding observation is that accuracy remains consistently high up to approximately $1$6 km and degrades beyond $1$7 km because of lower-quality chips (Sami et al., 2023).

For the unlabeled VIS target domain, the transductive classifier reaches $1$8 overall test accuracy, also reported elsewhere as $1$9. Lower performance is specifically noted for MT-LB, Sport vehicle, and ZSU23-4. The paper further reports a Fréchet Inception Distance of $207$00 for VIS images generated from MWIR, which the authors describe as indicating successful generation quality. Semi-supervised fine-tuning improves VIS accuracy to $207$01 with $207$02 labeled VIS, $207$03 with $207$04, and $207$05 with $207$06 (Sami et al., 2023).

In the detection-oriented YOLOatr evaluation on the DSIAC MWIR dataset, correlated testing $207$07 yields approximately $207$08 mAP for YOLOv5s and approximately $207$09 mAP for YOLOatr. Under decorrelated testing $207$10, where the model is trained on lower ranges and tested on higher ranges, the reported mAP values are approximately $207$11 for YOLOv5s and approximately $207$12 for YOLOatr. Per-class decorrelated YOLOatr [email protected] is reported as approximately $207$13 for T-72 tank, $207$14 for SUV, $207$15 for BTR70 or BRDM2 as named elsewhere, and $207$16 for pickup truck. Aggregate decorrelated precision and recall are approximately $207$17 and $207$18, respectively, and inference speed is reported as $207$19 fps with per-image latency of approximately $207$20 ms on a Tesla P100 (Safdar et al., 15 Jul 2025).

Several limitations are explicit in these studies. The MWIR$207$21VIS transfer results highlight a significant domain gap despite strong MWIR source performance, and the authors suggest that CycleGAN’s implicit bijection and cycle-consistency constraints may be overly strict for this application; they recommend exploring unpaired translation methods that relax bijection, including CUT and Spatially-Correlative Loss (Sami et al., 2023). The YOLOatr paper, meanwhile, documents range-induced degradation under decorrelated testing, attributing it to few-pixel targets, clutter changes, and loss of structural detail at higher ranges (Safdar et al., 15 Jul 2025).

The benchmark description itself also contains unresolved inconsistencies in one source. The YOLOatr paper states that Table 1 implies approximately $207$22 frames and approximately $207$23 videos, while another row lists $207$24 ARF files; it also alternates between “10 vehicles” and “13 different tactical and civilian vehicles,” without reconciliation (Safdar et al., 15 Jul 2025). For replication-oriented work, this suggests that exact dataset accounting should be documented explicitly rather than assumed from secondary summaries.