MIQD-2.5M: Machine-Centric Image Quality Database

• The paper introduces MIQD-2.5M, a novel benchmark that defines machine-centric image quality through consistency, accuracy, and composite metrics aggregated over 75 models.
• MIQD-2.5M is a large-scale database with 10,000 original images and 2.5M degraded samples across classification, detection, and segmentation tasks, enabling precise performance assessment.
• It employs a region-aware degradation taxonomy with 10 distortion types and 5 severity levels, highlighting the importance of spatially-sensitive quality evaluation for machine vision.

Searching arXiv for the cited MIQD/MIQA papers and precursors. Fetching metadata from arXiv for the specified papers. {"cmd":"python - << 'PY'\nimport urllib.request, xml.etree.ElementTree as ET\nids=['(Cao et al., 2022)','(Li et al., 13 Mar 2025)','(Wang et al., 27 Aug 2025)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n data=urllib.request.urlopen(url, timeout=20).read()\n root=ET.fromstring(data)\n ns={'a':'http://www.w3.org/2005/Atom'}\n entry=root.find('a:entry',ns)\n title=entry.find('a:title',ns).text.strip().replace('\n',' ')\n published=entry.find('a:published',ns).text.strip()\n print(aid, published, title)\nPY"}/Subthreshold്ടool output to=commentary 天天中彩票网_code (Cao et al., 2022) 2022-07-29T17:58:14Z Image Quality Assessment: Integrating Model-Centric and Data-Centric Approaches (Li et al., 13 Mar 2025) 2025-03-13T07:18:10Z Image Quality Assessment: From Human to Machine Preference (Wang et al., 27 Aug 2025) 2025-08-27T17:50:22Z Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models

Machine-Centric Image Quality Database (MIQD-2.5M) is a machine-centric image quality database introduced within the machine-centric image quality assessment (MIQA) paradigm, in which image quality is defined by the impact of degradations on machine vision systems (MVS) rather than by human perceptual judgment. The database comprises 10,000 original images and 2,500,000 degraded samples, spans 3 vision tasks and 75 vision models, and provides labels derived from consistency, accuracy, and their composite under distortion (Wang et al., 27 Aug 2025). In the broader evolution of IQA, MIQD-2.5M follows earlier work that argued for integrating model-centric failure analysis into dataset construction (Cao et al., 2022) and complements the machine-preference perspective developed in the Machine Preference Database (MPD), which formalized quality in terms of downstream task behavior across LMMs and CV models (Li et al., 13 Mar 2025).

1. Conceptual basis of machine-centric image quality

Conventional IQA was largely developed for the human visual system (HVS), whereas MIQD-2.5M is grounded in the requirements of machine vision systems (MVS). The motivating claim is that these two consumers of images do not perceive quality in the same way: HVS-based metrics such as PSNR, SSIM, LPIPS, DISTS, HyperIQA, MANIQA estimate human perceptual fidelity, but machine-oriented quality must reflect whether downstream predictions remain stable and correct under degradation (Wang et al., 27 Aug 2025).

Within this framing, MIQA defines image quality through downstream machine behavior. The 2025 MIQA paper explicitly emphasizes that quality for machines is not only “fidelity”; it also includes whether predictions stay stable and/or correct under distortion. This is operationalized in MIQD-2.5M through separate consistency and accuracy labels, rather than a single human-opinion score (Wang et al., 27 Aug 2025).

This machine-centric perspective is aligned with two adjacent developments. First, earlier work on integrating model-centric and data-centric IQA argued that dataset construction should be guided by current model failures, using a sampling-worthiness framework over a large unlabeled pool $\mathcal{U}$ to select samples that are both difficult and diverse (Cao et al., 2022). Second, the MPD work argued that IQA should move “from human to machine preference,” defining quality by how distortions alter downstream task outputs across tasks such as YoN, MCQ, VQA, CAP, SEG, DET, and RET (Li et al., 13 Mar 2025). Taken together, these studies place MIQD-2.5M within a research trajectory that treats machine utility, not human opinion, as the target variable.

2. Dataset composition and source images

MIQD-2.5M is constructed from 10,000 original images drawn from two standard benchmarks: the ImageNet validation set for image classification and the MS COCO validation set for object detection and instance segmentation. The database uses 5,000 ImageNet originals and 5,000 COCO originals (Wang et al., 27 Aug 2025).

For the ImageNet portion, the paper does not use all validation images. Instead, each image is scored by two pretrained HVS-based IQA models, HyperIQA and MANIQA, and the top 5 highest-rated images per class are selected. This yields a high-quality, class-balanced classification subset. For the COCO portion, all 5,000 validation images are used because the validation set is relatively smaller and already structured for detection and segmentation (Wang et al., 27 Aug 2025).

The resulting organization is summarized below.

Task	Source	Images
Image classification	ImageNet	5,000 originals, 1,250,000 degraded images
Object detection + instance segmentation	MS COCO	5,000 originals, 1,250,000 degraded images
Total	Combined	10,000 originals, 2,500,000 degraded images

This construction reflects a deliberate task-oriented design. Classification, detection, and instance segmentation are treated as three representative machine-vision tasks, and the database is explicitly built to support machine-centric learning and benchmarking rather than subjective human annotation (Wang et al., 27 Aug 2025).

3. Degradation taxonomy and region-aware organization

MIQD-2.5M includes 10 degradation types, 5 severity levels, and 3 spatial degradation patterns. The degradation types are grouped into four categories:

• Digital distortions: Contrast, Pixelate, JPEG compression
• Blur distortions: Motion blur, Defocus blur, Glass blur
• Environmental conditions: Fog, Snow, Darkness
• Noise distortions: Gaussian noise

Each degradation type is applied at 5 severity levels, from mild to severe (Wang et al., 27 Aug 2025).

A distinguishing feature of MIQD-2.5M is its use of region-aware degradations. Each degraded image is generated in one of three region-based forms:

• UD — Uniform Distortion: degradation applied equally to the whole image
• ROI-DD — Region-of-Interest Dominated Distortion: stronger distortion in task-relevant regions
• BG-DD — Background Dominated Distortion: stronger distortion in the background

For each raw image, the paper states that the database creates 250 degraded versions per original image by combining 10 distortion types with 5 severity combinations for ROI and background and uniform distortion (Wang et al., 27 Aug 2025).

This organization is important because prior IQA datasets usually assume spatially uniform distortion. MIQD-2.5M instead encodes the proposition that machine quality depends not only on how strong a corruption is, but also on where it occurs. The paper reports that UD is easiest to model, ROI-DD is harder, and BG-DD is hardest; in classification, BG-DD is especially difficult because background corruption weakens contextual reasoning (Wang et al., 27 Aug 2025). A plausible implication is that machine-centric quality estimation cannot be reduced to global fidelity alone, particularly for tasks that exploit object-context interactions.

4. Labeling scheme: consistency, accuracy, and composite quality

The central annotation principle of MIQD-2.5M is that labels are generated from machine responses. Let $x$ denote an original image, $x'$ a degraded image, and $\mathcal{M} = \{m_1, \dots, m_n\}$ a set of vision models with weights $\alpha_i$ such that $\sum_i \alpha_i = 1$ (Wang et al., 27 Aug 2025).

For one model, the consistency score is defined as

$C_{m_i}(x, x') = \mathbf{M}(m_i(x'), m_i(x)),$

and the aggregated consistency is

$$\mathbb{E}_C(x, x') = \sum_{i=1}^{n} \alpha_i C_{m_i}(x, x').$$

For one model, the accuracy score is defined against ground truth $g^*(x)$ as

$A_{m_i}(x, x') = \mathbf{M}(m_i(x'), g^*(x)),$

and the aggregated accuracy is

$x$0

The paper also defines a composite score

$x$1

with

$x$2

The special cases are $x$3 for pure consistency and $x$4 for pure accuracy (Wang et al., 27 Aug 2025).

The task-specific agreement function $x$5 depends on the vision task. For classification,

$x$6

For detection, the paper uses greedy one-to-one matching under IoU threshold $x$7:

$x$8

Precision and recall are computed over matched boxes,

$x$9

$x'$0

and detection performance is summarized by mAP. Instance segmentation uses the same framework, but matching is mask-based instead of box-based (Wang et al., 27 Aug 2025).

The label-generation process aggregates these quantities over 75 models. Label stability is validated by randomly splitting models into 80% / 20%, generating labels independently, and repeating this 100 times. The reported cross-model label consistency is high: for example, in classification, composite labels achieve SRCC 0.914 and PLCC 0.952; in object detection, composite labels achieve SRCC 0.959 and PLCC 0.960; and in instance segmentation, composite labels achieve SRCC 0.959 and PLCC 0.960 (Wang et al., 27 Aug 2025).

5. Model coverage, benchmarking protocol, and RA-MIQA

MIQD-2.5M spans 75 models across the three supported tasks. The classification portion includes CNNs, Vision Transformers, hybrid models, and other modern architectures, with examples such as VGG19, ResNet-34/50/101, Inception-v4, SE-ResNet / SE-ResNeXt, EfficientNet-B2/B4/B6/B7, ConvNeXt variants, ViT, Swin, DeiT, CaiT, MViTv2, CoaT, CoAtNet, LeViT, Visformer, MLP-Mixer, gMLP, MambaOut. Detection models include Faster R-CNN, Cascade R-CNN, CenterNet, Sparse R-CNN, YOLOX / YOLOv9 / YOLOv10 / YOLOv11 variants, PVT, Swin, ViTDet, DINO, DDQ-DETR, CO-DETR, DETR, Deformable DETR. Segmentation models include Mask R-CNN, Cascade Mask R-CNN, PointRend, YOLO-seg variants, RTMDet-X, ConvNeXt-V2, Swin-based segmenters, ViTDet, and Mask2Former (Wang et al., 27 Aug 2025).

The associated benchmark uses a 4:1 split with 80% training and 20% validation, with the constraint that original images and their degraded versions do not leak across partitions. Evaluation uses SRCC, PLCC, KRCC, and RMSE; score alignment for PLCC and RMSE uses the five-parameter logistic function

$x'$1

The reported training details are PyTorch, NVIDIA RTX 3090, 5 epochs, batch size 256, AdamW, learning rate $x'$2, weight decay $x'$3, cosine annealing, 1-epoch warmup, resize shorter side to 288, random crop $x'$4 for training, center crop $x'$5 for validation, horizontal and vertical flips with probability 0.5, and MSE loss (Wang et al., 27 Aug 2025).

The paper introduces RA-MIQA, a region-aware machine-centric image quality assessment model, with three components: a frozen region encoder, a Transformer encoder, and a regression head. It maps an image to a scalar quality score,

$x'$6

Its region-aware mechanism uses patch embeddings $x'$7, a frozen pretrained region encoder $x'$8 to obtain $x'$9, average pooling, projection into the transformer embedding space,

$\mathcal{M} = \{m_1, \dots, m_n\}$0

and a token sequence

$\mathcal{M} = \{m_1, \dots, m_n\}$1

The final prediction is

$\mathcal{M} = \{m_1, \dots, m_n\}$2

The paper describes the head as a two-layer MLP with learnable parameters $\mathcal{M} = \{m_1, \dots, m_n\}$3 (Wang et al., 27 Aug 2025).

Empirically, the best HVS-based metric on composite score prediction is usually DISTS, but it remains substantially below machine-centric methods. For classification, DISTS achieves SRCC 0.3878 while RA-MIQA achieves SRCC 0.7003; for object detection, 0.5266 versus 0.8125; and for instance segmentation, 0.5363 versus 0.8188. Compared with ViT-small, RA-MIQA improves composite SRCC by +16.76% on classification, +1.66% on detection, and +2.76% on segmentation. The region encoder ablation further reports gains of +30.61% SRCC / +23.29% PLCC for classification, +4.22% / +3.79% for detection, and +4.56% / +3.99% for segmentation when the region encoder is added to a ResNet-18 baseline (Wang et al., 27 Aug 2025).

6. Relation to MPD and to failure-driven dataset construction

MIQD-2.5M belongs to a broader transition from human-centric IQA to machine-centric data construction. A related 2025 database, the Machine Preference Database (MPD), formalizes machine preference through 7 downstream tasks, 15 LMMs + 15 CV models = 30 machine subjects, 30 corruption scenarios, 5 corruption strengths, 1,000 reference images, 30,000 reference/distorted image pairs, and 2.25 million fine-grained annotations (Li et al., 13 Mar 2025).

The two resources differ in emphasis. MPD defines machine preference using tasks such as YoN, MCQ, VQA, CAP, SEG, DET, and RET, and aggregates these into a machine MOS. MIQD-2.5M, by contrast, is centered on classification, object detection, and instance segmentation, and it labels image quality through consistency, accuracy, and composite machine-response signals across 75 vision models (Li et al., 13 Mar 2025). MPD also spans NSIs, SCIs, and AIGIs, whereas MIQD-2.5M is built from ImageNet and COCO with region-aware degradations (Wang et al., 27 Aug 2025).

An even earlier precursor is the work on integrating model-centric and data-centric IQA. That paper did not introduce MIQD-2.5M, but it explicitly motivated machine-centric construction of next-generation IQA databases by proposing a closed-loop framework: start from a large unlabeled image pool $\mathcal{M} = \{m_1, \dots, m_n\}$4, use a model-centric IQA system $\mathcal{M} = \{m_1, \dots, m_n\}$5 to estimate which images are difficult and diverse, select a subset $\mathcal{M} = \{m_1, \dots, m_n\}$6, conduct subjective testing, improve the model, and repeat (Cao et al., 2022). Its subset-selection objective,

$\mathcal{M} = \{m_1, \dots, m_n\}$7

together with the failure predictor $\mathcal{M} = \{m_1, \dots, m_n\}$8 and the semantic diversity term based on 1,000-dimensional logits of VGGNet, established the methodological principle that datasets should be built to expose model failure modes rather than sampled without machine feedback (Cao et al., 2022). This suggests that MIQD-2.5M can be understood not only as a benchmark, but also as part of a broader program of failure-driven data design.

7. Empirical implications and open technical issues

The main empirical implication of MIQD-2.5M is that HVS-based IQA is not a reliable proxy for MVS quality. The MIQA paper reports that HVS-based metrics are inadequate for MVS quality prediction, while even specialized MIQA models still struggle with background degradations, accuracy-oriented estimation, and subtle distortions (Wang et al., 27 Aug 2025). This conclusion is consistent with MPD, which found that mainstream IQA methods that work well for human perception do not accurately model machine preference, and that in Mild Distortion, no method even beats PSNR in some cases (Li et al., 13 Mar 2025).

The papers also indicate that machine sensitivity is task-dependent. In MPD, inter-task correlation is relatively low, especially between VQA and specialized CV tasks, and the overall MOS correlates better with each dimension than the dimensions correlate with each other. MIQD-2.5M likewise reports limited but informative cross-task generalization: models trained on detection and tested on segmentation generalize relatively well, whereas transfer to classification is weaker (Li et al., 13 Mar 2025). A plausible implication is that machine-centric IQA may require task-specific or task-conditional modeling rather than a universal scalar notion inherited from human perceptual IQA.

The practical consequences identified for MIQD-2.5M include estimating how badly a distortion will affect downstream machine perception, monitoring image quality before inference, and guiding camera calibration, exposure control, compression settings, denoising/restoration decisions, and transmission quality control. The same work also suggests future MIQA models should encode region sensitivity explicitly, distinguish consistency from accuracy, model task-specific quality perception, handle subtle and background-only degradations better, and support cross-task adaptation (Wang et al., 27 Aug 2025).

In that sense, MIQD-2.5M marks a shift in the ontology of image quality itself. Instead of treating quality as a human opinion score or perceptual-fidelity estimate, it treats quality as the degree to which an image preserves machine behavior under controlled degradation. For machine vision research, this reframing is the database’s defining contribution (Wang et al., 27 Aug 2025).