DroneSR Dataset for Thermal Image Super-Resolution

• DroneSR is a few-shot thermal infrared dataset designed to test super-resolution models under scarce data conditions and overfitting risks.
• It integrates multi-source imagery from aerial and ground acquisitions to simulate diverse real-world drone scenarios.
• Evaluated at 3×, 4×, and 6× upscale factors, DroneSR benchmarks performance using metrics such as PSNR, SSIM, LPIPS, FSIM, and MS-SSIM.

Searching arXiv for the specified paper and related DroneSR/DSR works. {"query":"ti:\"DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective\" OR id:(Weng et al., 2 Sep 2025)","max_results":5} {"query":"DroneSR dataset thermal super-resolution (Weng et al., 2 Sep 2025)", "max_results": 10} arXiv search: DroneSR (Weng et al., 2 Sep 2025) related datasets and prior drone SR benchmarks DroneSR is the drone-captured thermal imaging super-resolution benchmark introduced by Zhipeng Weng, Xiaopeng Liu, Ce Liu, Xingyuan Guo, Yukai Shi, and Liang Lin in “DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective” (Weng et al., 2 Sep 2025). It targets thermal infrared single-image super-resolution under few-shot training and was constructed to expose a specific failure mode of large-scale generative architectures: severe overfitting when training data consist of scarce, drone-based infrared imagery. The benchmark is explicitly multi-source, combining aerial infrared, ground infrared, and additional outdoor infrared imagery, and it is used to evaluate super-resolution at $\times 3$, $\times 4$, and $\times 6$ upscaling factors.

1. Problem setting and benchmark rationale

DroneSR is organized around a narrow but technically consequential problem: thermal infrared single-image super-resolution in the low-data regime. Its stated motivation is that large-scale generative architectures, including diffusion models, frequently overfit under few-shot drone-based thermal imaging conditions. The dataset is therefore not merely a corpus of thermal images; it is a benchmark constructed to foreground aerial infrared data scarcity, multi-source sensor diversity, and the robustness of super-resolution models under realistic drone scenarios.

Its defining characteristic is the combination of a drone-based perspective with multi-source heterogeneity. Prior infrared super-resolution datasets are described as focusing on a single sensor or on larger-scale curated sources, whereas DroneSR is explicitly multi-source, few-shot, and intended to stress multi-sensor variability together with complex degradations common in real drone operations. The benchmark centers on drone-based imagery while also mixing ground infrared data to increase diversity and domain heterogeneity.

A plausible implication is that DroneSR is designed less as a maximal-scale training resource than as a stress test for model inductive bias and regularization. In that sense, the benchmark operationalizes overfitting as a first-class evaluation concern rather than treating it as a secondary training artifact.

2. Data sources, modality, and split structure

The modality is thermal infrared imagery. The paper does not specify the sensor types or camera models, and the spectral band is also not specified. One cited source within the benchmark, the IRay Infrared Image Platform, provides $1280 \times 1024$ high-resolution thermal images featuring dynamic urban scenes; other images are collected from online drone-captured aerial infrared datasets and outdoor infrared datasets, but exact resolutions for those sources are not specified (Weng et al., 2 Sep 2025).

DroneSR is described as a multi-source composition with three components: aerial infrared sources collected from the internet, ground infrared images crawled from the IRay Infrared Image Platform, and additional outdoor infrared datasets from the internet. The authors emphasize “complex and diversified infrared imagery” across aerial and ground platforms, including dynamic urban scenes. Specific geographic regions, rural or maritime domains, time-of-day, weather conditions, flight altitudes, and distances are not specified.

The benchmark follows a deliberately limited few-shot split.

Split	Composition	Count
Train	150 aerial + 50 ground	200
Validation	30 aerial + 20 ground	50
Test	Total for metric evaluation	50

The sampling strategy is random. No detection categories, bounding boxes, or related annotation structures are specified for the reported benchmark, despite the abstract’s phrase “a multi source drone-based infrared image benchmark dataset for detection.” The reported splits and experiments are super-resolution-focused rather than detection-focused.

3. High-resolution data, synthetic low-resolution construction, and degradation design

DroneSR provides high-resolution images, while low-resolution counterparts are generated synthetically during experiments. For evaluation and visualization, high-resolution images are downsampled by a factor $s \in \{3,4,6\}$ to form low-resolution inputs. For example, one visualization protocol obtains the low-resolution image by downsampling the original image by a factor of four. Because low-resolution images are derived from the high-resolution source through known degradations, registration and external alignment are not discussed.

The benchmark is tightly coupled to Gaussian Quantization Representation Learning (GQRL), which is applied during training to produce a low-quality input $I_{\mathrm{quant}}$ for a two-stage super-resolution model. The Gaussian quantization uses the probability density

$$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$

where $x$ ranges in $[0,255]$, $\mu$ is the mean, and $\times 4$0 is the standard deviation. The interval count $\times 4$1 is sampled from a truncated and stepped Gaussian $\times 4$2, clipped to $\times 4$3, and snapped to a step of $\times 4$4. Boundaries $\times 4$5 partition the per-channel range $\times 4$6, and pixels inside each interval are reassigned to representative values, producing $\times 4$7.

Algorithm 1 lists three quantization strategies—“middle,” “random,” and “zero”—while the main text describes random representative value selection within intervals. The experimental parameterization uses $\times 4$8, $\times 4$9, batch size $\times 6$0, and diffusion steps often set to $\times 6$1 as a balance during training.

Bit-depth is not formally specified. The processing pipeline assumes pixel intensities in $\times 6$2, which is consistent with 8-bit thermal imagery, but the paper does not explicitly state the bit-depth. Normalization and augmentation beyond quantization and downsampling are likewise not specified. The paper also highlights prevalent noise, variable image quality, and motion blur for fast-moving targets, but it does not provide quantified sensor noise or compression parameters.

4. Evaluation protocol and benchmarked performance

DroneSR evaluates super-resolution at $\times 6$3, $\times 6$4, and $\times 6$5. The reported metrics are PSNR, SSIM, LPIPS, FSIM, and MS-SSIM on the 50-image test set (Weng et al., 2 Sep 2025). For completeness, the standard definitions included alongside the benchmark description are

$\times 6$6

and

$\times 6$7

The paper uses LPIPS, FSIM, and MS-SSIM as additional metrics but does not print their mathematical forms.

At $\times 6$8, which is the paper’s primary comparison setting, the evaluated methods are Real-ESRGAN, LIRSN, DiffBIR, StableSR, Resshift, and the proposed method. Their reported test-set values are:

• Real-ESRGAN: PSNR $\times 6$9; SSIM $1280 \times 1024$0; LPIPS $1280 \times 1024$1; FSIM $1280 \times 1024$2; MS-SSIM $1280 \times 1024$3.
• LIRSN: PSNR $1280 \times 1024$4; SSIM $1280 \times 1024$5; LPIPS $1280 \times 1024$6; FSIM $1280 \times 1024$7; MS-SSIM $1280 \times 1024$8.
• DiffBIR: PSNR $1280 \times 1024$9; SSIM $s \in \{3,4,6\}$0; LPIPS $s \in \{3,4,6\}$1; FSIM $s \in \{3,4,6\}$2; MS-SSIM $s \in \{3,4,6\}$3.
• StableSR: PSNR $s \in \{3,4,6\}$4; SSIM $s \in \{3,4,6\}$5; LPIPS $s \in \{3,4,6\}$6; FSIM $s \in \{3,4,6\}$7; MS-SSIM $s \in \{3,4,6\}$8.
• Resshift: PSNR $s \in \{3,4,6\}$9; SSIM $I_{\mathrm{quant}}$0; LPIPS $I_{\mathrm{quant}}$1; FSIM $I_{\mathrm{quant}}$2; MS-SSIM $I_{\mathrm{quant}}$3.
• Ours: PSNR $I_{\mathrm{quant}}$4; SSIM $I_{\mathrm{quant}}$5; LPIPS $I_{\mathrm{quant}}$6; FSIM $I_{\mathrm{quant}}$7; MS-SSIM $I_{\mathrm{quant}}$8.

Across scales, the paper reports the following summary for its own method and the principal exceptions in LPIPS:

Scale	Reported values for the proposed method	Notable metric exception
$I_{\mathrm{quant}}$9	PSNR $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$0; SSIM $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$1; LPIPS $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$2; FSIM $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$3; MS-SSIM $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$4	Real-ESRGAN has the best LPIPS: $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$5
$$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$6	PSNR $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$7; SSIM $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$8; LPIPS $$p_{\mathrm{gaussian}}(x)=\frac{1}{\sqrt{2\pi}\,\sigma}\,\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right),$$9; FSIM $x$0; MS-SSIM $x$1	Primary comparison setting
$x$2	PSNR $x$3; SSIM $x$4; LPIPS $x$5; FSIM $x$6; MS-SSIM $x$7	DiffBIR has the best LPIPS: $x$8

These results are used in the paper to argue that DroneSR is difficult for large super-resolution models trained under scarce thermal-data conditions, especially when robustness to overfitting is inadequate.

5. Overfitting as a benchmarked phenomenon

A central empirical claim of DroneSR is that few-shot thermal infrared super-resolution exposes overfitting in large generative architectures. The benchmark is used to show that DiffBIR exhibits overfitting under the DroneSR regime, with validation loss and PSNR degradation after approximately $x$9 steps, whereas the proposed method maintains convergence (Weng et al., 2 Sep 2025). Validation loss and PSNR curves are shown in Figure 1, although detailed monitoring formulas are not provided.

The benchmark is also the experimental context for Gaussian Quantization Representation Learning. The paper states that Gaussian-guided quantization creates diverse degradations from scarce high-resolution infrared data, smooths pixel distributions, reduces noise, preserves edges and structure, and thereby mitigates overfitting in large generative super-resolution models. In this formulation, the dataset is inseparable from the method’s intended regularization effect: scarcity and heterogeneity are not nuisances to be abstracted away, but the very conditions under which the approach is tested.

The latent diffusion training objective is written in the paper as

$[0,255]$0

with $[0,255]$1 via a VAE encoder, and

$[0,255]$2

The denoising objective is

$[0,255]$3

and the coarse reconstruction loss is

$[0,255]$4

The abstract additionally mentions an effective monitoring mechanism for tracking large architectures during training, but its implementation details are not specified in the main text. This leaves the overfitting claim well illustrated empirically but only partially formalized procedurally.

6. Release status, reproducibility, and intended usage

The paper states that “The code and DroneSR dataset will be available at: https://github.com/wengzp1/GARLSR” (Weng et al., 2 Sep 2025). It does not provide a separate dataset URL, versioning scheme, or license in the text.

The reported experiments are conducted on an NVIDIA RTX 3090 GPU. The stated training configurations include batch size $[0,255]$5, training steps of $[0,255]$6 for the coarse network, $[0,255]$7 for diffusion, $[0,255]$8 for the denoiser, and diffusion steps often set to $[0,255]$9. File structure, naming conventions, and loader scripts are not described.

The practical benchmark protocol is correspondingly narrow. To reproduce the reported setting, one constructs low-resolution–high-resolution pairs by downsampling the high-resolution images by $\mu$0, optionally applies Gaussian quantization to generate $\mu$1 for training, and evaluates PSNR, SSIM, LPIPS, FSIM, and MS-SSIM on the 50-image test split. A plausible implication is that reproducibility depends more on faithfully reconstructing the degradation and training pipeline than on any elaborate annotation schema, because the benchmark is based on high-resolution images plus synthetic low-resolution generation rather than on pre-packaged native low-resolution captures.

7. Relation to earlier drone super-resolution datasets, naming ambiguity, and limitations

The name “DroneSR” is not unique in the drone super-resolution literature. “DSR: Towards Drone Image Super-Resolution” introduced the Drone Super-Resolution dataset, also referred to as DSR and in some discussions as DroneSR, for top-view RGB drone imagery across varying altitudes (Lin et al., 2022). The EPFL IVRL thesis “Towards Robust Drone Vision in the Wild” by Xiaoyu Lin also uses the official name DroneSR for a real-world drone-vision super-resolution dataset with paired low-resolution and high-resolution captures, altitude metadata, and an effective scale factor of $\mu$2 (Lin, 2022).

Those earlier datasets differ substantially from the 2025 DroneSR benchmark. The DSR line is based on native paired captures from two cameras onboard a DJI Mavic 3, focuses on nadir or top-view RGB and RAW imagery, records altitude explicitly, and studies domain gaps across the altitude set $\mu$3 m. By contrast, DroneSR in the 2025 sense is thermal infrared, multi-source across aerial and ground platforms, few-shot by design, and generates low-resolution inputs synthetically from high-resolution thermal images rather than relying on native low-resolution–high-resolution optical pairing. The distinction is therefore not merely nominal; it marks a shift from altitude-conditioned RGB drone super-resolution to few-shot thermal infrared robustness and overfitting analysis.

Two points are especially easy to misconstrue. First, although the abstract mentions a benchmark “for detection,” no detection annotations or categories are specified, and the reported benchmark is super-resolution-centric. Second, while DroneSR is described as drone-based, it intentionally mixes ground infrared imagery to increase diversity and domain heterogeneity rather than restricting itself to purely aerial capture.

The benchmark’s stated limitations are also clear. Motion blur in fast-moving targets can cause information loss and underperforming super-resolution outputs. Dataset scale is intentionally few-shot, and broader coverage or annotations such as detection labels are not part of the reported benchmark. Ethical considerations are not discussed. The authors identify motion-blur handling and related degradations as a direction for future work (Weng et al., 2 Sep 2025).