Dynamic-ISTD: Infrared Target Benchmark

• The benchmark introduces a controlled evaluation substrate that models cross-domain shifts, UAV motion blur, and signal-dependent noise perturbations.
• It comprises 206 standardized infrared images from diverse aerial scenes with pixel-level annotations and a balanced 50/50 train-test split.
• The construction pipeline employs wavelet-guided synthesis, Poisson blending, and dual-branch noise invariance training to ensure robust performance under dynamic degradations.

Searching arXiv for the specified benchmark paper and closely related context. {"query":"arXiv (Li et al., 14 Oct 2025) Ivan-ISTD Dynamic-ISTD Benchmark infrared small target detection", "max_results": 5} Searching arXiv for cross-domain infrared small target detection context. The Dynamic-ISTD Benchmark is a purpose-built, cross-domain infrared small-target detection dataset that simulates both background distribution shifts and dynamic, heteroscedastic noise degradations encountered in real UAV flight scenarios. It was introduced in "Ivan-ISTD: Rethinking Cross-domain Heteroscedastic Noise Perturbations in Infrared Small Target Detection" (Li et al., 14 Oct 2025). Within that framework, the benchmark functions as a controlled evaluation substrate for studying robustness under simultaneous domain shift, motion blur, and signal-dependent noise, rather than as a static collection of infrared frames alone.

1. Dataset definition and composition

Dynamic-ISTD contains 206 infrared images, and all frames are standardized to $640\times512$ pixels in grayscale (Li et al., 14 Oct 2025). The images are drawn from multiple aerial scenes, including urban roads, farmland, coastal or maritime environments, and forested areas. Three semantic target classes appear across all scenes, with the paper giving vehicles, vessels, and human figures as examples; these targets are annotated at the pixel level.

The benchmark uses an even $50/50$ split into training and test sets, with 103 images in each partition. Within the training set, 70% of images remain unaltered as “natural” samples and 30% receive controlled motion-blur degradation to simulate in-flight motion. The test set is held out and degraded more aggressively. The paper further notes that, if a separate validation split is required, one may sub-sample approximately 10% of the training partition for hyperparameter tuning, leaving approximately 93 frames for learning.

Aspect	Specification	Note
Total images	206	Infrared, grayscale
Resolution	$640\times512$	Standardized across all frames
Split	103 train / 103 test	Even 50/50 partition
Training degradation	70% natural, 30% motion blur	Controlled in-flight simulation
Test degradation	Multi-angle blur + composite noise	Held-out, more aggressive

This composition makes the benchmark explicitly cross-domain. The benchmark is not limited to scene variation alone; its design couples background shift with dynamic degradations, so the data distribution encountered at test time is intentionally misaligned with the training distribution.

2. Degradation model

The benchmark models two principal degradation axes: spatial blur for UAV motion and realistic heteroscedastic noise perturbations (Li et al., 14 Oct 2025). In the training set, 30% of images receive uniform linear motion blur with blur length $l=5$ pixels and blur angle $\theta=45^\circ$. In the test set, each image is corrupted by one of several discrete blur angles,

$$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$

using the same blur length $l=5$. If $x$ denotes the original image, the blurred image $x^b$ is defined by

$x^b = x * k_{l,\theta},$

where $50/50$0 is the one-dimensional motion-blur kernel with length $50/50$1 and orientation $50/50$2.

The noise model includes five synthetic noise types—Gaussian, salt-&-pepper, speckle, uniform, and Poisson—as well as real-domain noise extracted from held-out target images. All test-set images receive composite synthetic noise at density $50/50$3, with an optional real-noise overlay. To represent signal-dependent corruption, the paper gives the heteroscedastic model

$50/50$4

where $50/50$5 may be estimated from local intensity, for example as a fraction of the pixel value or of the background variance.

Real-domain noise extraction is based on target-domain background imagery without annotated targets. A sliding window of size $50/50$6 is applied to each background image, and local mean $50/50$7 and variance $50/50$8 are computed. Only patches satisfying $50/50$9 and $640\times512$0 are retained, to avoid saturated or excessively textured regions. These patches are then upsampled back to full image size to form a noise library $640\times512$1. Mixed noisy samples are generated by

$640\times512$2

where $640\times512$3 is sampled randomly each iteration, and the paper reports that $640\times512$4 was found empirically optimal.

A plausible implication is that Dynamic-ISTD is designed to probe both additive corruption robustness and robustness to mismatch in the underlying background statistics. That interpretation follows directly from the benchmark’s joint emphasis on synthetic composite noise, real-domain noise extraction, and cross-domain background synthesis.

3. Construction pipeline

The benchmark construction is organized into two stages: Wavelet-guided Cross-domain Sample Synthesis and Real-domain Noise Invariance Learning (Li et al., 14 Oct 2025). The first stage produces cross-domain synthetic samples aligned with target-domain backgrounds; the second stage builds a training process in which real-noise perturbations are explicitly incorporated.

Stage I: Wavelet-guided Cross-domain Sample Synthesis

The first subprocedure is Background Region Detection. Each target-domain frame $640\times512$5 is decomposed with a one-level Haar wavelet into a low-frequency component $640\times512$6 and three high-frequency bands $640\times512$7. Each high-frequency band is edge-aware filtered as

$640\times512$8

with $640\times512$9. A denoised feature map is then reconstructed: $l=5$0

The reconstructed feature map is tiled into $l=5$1 non-overlapping blocks of size $l=5$2. For each block $l=5$3, the score

$l=5$4

is computed, and all blocks satisfying $l=5$5 are selected as “pure” background. These selected blocks are upsampled back to full $l=5$6 resolution via bilinear interpolation.

The second subprocedure is Dual-Indicator Hard-Target Selection from the source set. The current detector $l=5$7 is run on each source image $l=5$8 to produce a soft mask $l=5$9. For each annotated target region, pixel accuracy and IoU are compared to thresholds $\theta=45^\circ$0 and $\theta=45^\circ$1. Source crops are retained as hard examples only when both $\theta=45^\circ$2 and $\theta=45^\circ$3.

The third subprocedure is SSIM-guided Poisson Re-generation. For each hard source patch $\theta=45^\circ$4, SSIM is computed against candidate target backgrounds $\theta=45^\circ$5, and the top-scoring location $\theta=45^\circ$6 is selected. Fusion is then carried out by Poisson blending: $\theta=45^\circ$7 The resulting synthetic frame is described as geometrically consistent with the target background and photometrically seamless.

Stage II: Real-domain Noise Invariance Learning

The second stage uses the real-noise library described above and samples noisy inputs for each batch according to

$\theta=45^\circ$8

Training employs a dual-branch network with shared weights. The main branch processes clean $\theta=45^\circ$9, while the auxiliary branch processes noisy $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$0.

Two losses are combined. The supervised multi-scale BCE loss is

$$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$1

and the self-supervised feature-consistency loss is

$$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$2

The total objective is

$$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$3

The paper states that pseudocode for Background Region Detection and re-generation is provided in Algorithms 1 and 2, while the noise-library extraction and mixing script is described as a straightforward sliding-window loop followed by interpolation.

4. Evaluation protocol and metrics

The prescribed training regime is 1,000 epochs with batch size 8, the Adam optimizer, and an initial learning rate of $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$4 cosine-annealed to $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$5 (Li et al., 14 Oct 2025). Cross-domain testing trains on the 50% training subset with limited-angle blur and evaluates on the held-out 50% test subset with multi-angle blur and composite noise. The paper’s experimental guidelines specify that training should always be performed on the training subset with blur angle fixed at $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$6, while testing should be performed on the held-out subset with multi-angle motion blur and 5% composite noise.

Evaluation uses per-pixel counts $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$7, $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$8, $$\theta \in \{25^\circ,45^\circ,90^\circ,135^\circ,150^\circ,\dots\},$$9, and $l=5$0 aggregated across the test frames. The reported metrics are Pixel Accuracy, Mean Intersection over Union, Normalized IoU, Detection Probability, False Alarm Rate, and F1-score. Pixel Accuracy is defined as

$l=5$1

Detection Probability is given as recall,

$l=5$2

and False Alarm Rate is

$l=5$3

The F1-score is

$l=5$4

The protocol requires reporting PixAcc, mIoU, nIoU, $l=5$5, $l=5$6, and F1 for direct comparison. For cross-dataset validation, the same protocol is to be followed when transferring between NUAA-SIRST and IRSTD-1K. This suggests that Dynamic-ISTD is intended to support both in-benchmark robustness assessment and out-of-domain transfer studies under consistent degradation rules.

5. Reproduction and experimental use

The paper frames Dynamic-ISTD as reproducible with a relatively explicit workflow (Li et al., 14 Oct 2025). Data preparation begins with collecting or downloading 206 infrared frames at $l=5$7, manually annotating all small targets at the pixel level, partitioning the data into 50% train and 50% test, applying 30% motion blur to the training set, and applying multi-angle blur plus composite noise to the test set.

Synthetic sample generation uses the Background Region Detection, SSIM, and Poisson blending pipeline to create cross-domain synthetic training images on the fly. In parallel, a real-noise library is built by sliding-window extraction over held-out background regions. Training then requires implementing the dual-branch network in a deep-learning framework, optimizing

$l=5$8

and using batch size 8, Adam, 1,000 epochs, and cosine learning-rate decay from $l=5$9 to $x$0.

The benchmark is also intended to be integrated into evaluation scripts. The usage guidelines specify including the Dynamic-ISTD test set in evaluation code, comparing pixel-wise and detection metrics to prior baselines, and ablating hyperparameters such as the mixing coefficient $x$1 and the background-selection threshold $x$2. The code for the Ivan-ISTD work is available at the repository identified in the paper: https://github.com/nanjin1/Ivan-ISTD.

6. Position within infrared small-target detection research

Dynamic-ISTD was introduced to address what the paper identifies as the dual challenges of cross-domain shift and heteroscedastic noise perturbations in infrared small-target detection (Li et al., 14 Oct 2025). In that sense, the benchmark is inseparable from the Ivan-ISTD framework: the dataset design, the wavelet-guided synthesis procedure, and the real-noise invariance training strategy are co-developed so that robustness can be evaluated under degradations meant to resemble real UAV deployment conditions.

A plausible misconception is to interpret Dynamic-ISTD as only a blur-and-noise augmentation suite. The construction described in the paper is broader. It includes multi-scene aerial imagery, pixel-level target annotations, explicit train/test asymmetry, wavelet-guided background alignment, hard-target selection, Poisson re-generation, and real-domain noise extraction. The benchmark therefore combines cross-domain composition with dynamic degradation modeling, rather than merely appending synthetic perturbations to a fixed corpus.

The paper also states that the versatility of the method is validated using other real-world datasets, and its experimental guidelines explicitly reference transfer evaluation between NUAA-SIRST and IRSTD-1K. This places Dynamic-ISTD in a broader methodological role: it serves as a controlled benchmark for robustness under distribution shift while remaining compatible with cross-dataset validation protocols commonly used in infrared small-target detection.