Image processing spans small object-recognition benchmarks, petabyte-scale satellite archives, biomedical repositories, and authenticity tests for manipulated media. The figures below show how different workloads are measured: images, labels, annotations, studies, storage volume, daily inflow, and model accuracy. Measurement periods and geography remain explicit because these statistics are snapshots from different releases and organizations.
Contents
- Benchmark collections
- Detection, segmentation, and annotation
- Biomedical imaging repositories
- Media forensics datasets
- Remote-sensing archives
- Medical image-processing evaluation
Benchmark collections
Large image-processing benchmarks provide the raw material for classification, detection, segmentation, and scene understanding. Their sizes differ substantially, and the split structure matters when interpreting a headline number.
Open Images V7 lists 9,011,219 images in its training split, 41,620 in validation, and 125,436 in test. These counts come from the Open Images V7 release snapshot described by Open Images V7 - Description. The three splits should be treated as separate release figures rather than as a single uniformly measured operational stream.
COCO reports 330,000 images, with more than 200,000 labeled. Its release snapshot also reports 80 object categories. The COCO - Common Objects in Context project describes the collection as an object- and context-oriented benchmark, so its image count and label count represent different dimensions of the dataset.
Places contains more than 10 million images and more than 400 unique scene categories, according to the Places: A 10 million Image Database for Scene Recognition release page. Its emphasis is scene recognition, which makes the category count a useful complement to the image-volume figure.
Caltech-256 is much smaller by image count: the Caltech 256 release lists 30,607 images organized into 256 object categories. The contrast with Places and Open Images illustrates why “dataset size” alone does not describe an image-processing task. Category breadth, annotation type, and split design also shape how a collection is used.
| Collection | Images reported | Category or split detail | Measurement period |
|---|---|---|---|
| Open Images V7 training | 9,011,219 | Training split | V7 release snapshot |
| Open Images V7 validation | 41,620 | Validation split | V7 release snapshot |
| Open Images V7 test | 125,436 | Test split | V7 release snapshot |
| COCO | 330,000 | 80 object categories; more than 200,000 labeled | Release snapshot |
| Places | More than 10 million | More than 400 scene categories | Places2 release snapshot |
| Caltech-256 | 30,607 | 256 object categories | Dataset release |
These figures are not directly interchangeable. Open Images split counts, COCO’s labeled subset, Places’ “more than” threshold, and Caltech-256’s fixed release count describe different collection designs.
Detection, segmentation, and annotation
Annotation volume can exceed image volume because one image may contain many boxes, masks, relationships, captions, points, or labels. Open Images V7 provides a particularly broad view of the processing workload. Its release snapshot reports 15,851,536 bounding boxes across 600 classes and 2,785,498 instance segmentations across 350 classes. For systems that identify both where an object is and which pixels belong to it, these are separate annotation tasks.
The same Open Images V7 snapshot reports 3,284,280 relationship annotations across 1,466 relationships and 675,155 localized-narrative annotations. It also lists 66,391,027 point-level annotations across 5,827 classes. These figures show how image processing can move beyond a single object label toward spatial, relational, and language-linked information.
Image-level labeling is larger still: Open Images V7 reports 61,404,966 image-level labels across 20,638 classes. That number is an annotation count, not a count of unique images or a measure of storage. The distinction matters when estimating data pipelines, review effort, or training inputs.
COCO reports 1.5 million annotated object instances, plus 91 stuff categories in addition to its object categories. It provides five captions per image and includes keypoints for 250,000 people, according to COCO - Common Objects in Context. These modalities support several image-processing problems in one collection: object localization, background or “stuff” recognition, image description, and human-pose analysis.
Together, the Open Images and COCO figures indicate why annotation counts should be reported alongside image counts. A benchmark with fewer images can still require substantial processing when it contains dense object instances, masks, points, captions, or keypoints.
Biomedical imaging repositories
Biomedical image processing is often described using cases, studies, series, images, collections, and storage volume at the same time. The NCI Imaging Data Commons (IDC) portal’s Data Release 24.0, dated April 27, 2026, lists 176 collections, 85,682 cases, 1,032,911 image series, and 99.27 TB of data volume. These are repository-level snapshot statistics from IDC, not a universal total for all biomedical imaging.
The current Imaging Data Commons overview says more than 95% of IDC data is covered by the permissive CC-BY license. It names nine contributing projects: TCGA, TCIA, CPTAC, CCDI, HTAN, LIDC, QIN, VHP, and NLST. Licensing and project provenance are operationally important because they affect how images can be reused, processed, and shared.
The Medical Imaging and Data Resource Center (MIDRC) reports more than 300,000 imaging studies collected and more than 135,000 released publicly on its 2026 program outcomes page. It also reports 27 algorithms in development and more than 450 data users. These numbers combine repository activity, algorithm development, and participation; they should not be read as equivalent measures of image volume. The figures are from Medical Imaging and Data Resource Center.
The Open Biomedical Imaging Archive provides another scale reference. As of September 2023, OBIA: An Open Biomedical Imaging Archive reported data for 937 individuals, 4,136 studies, 24,701 image series, and 1,938,309 images. It covered nine imaging modalities and 30 anatomical sites at that time. Because those counts are explicitly dated September 2023, they should not be treated as current repository totals without a newer measurement.
Media forensics datasets
Image authenticity work uses both pristine and manipulated media. The NIST MFC benchmark description from 2018 reports 35 million internet-sourced world-data images and 300,000 video clips in its world-data component. It also includes up to 176,000 pristine high-provenance images, 11,000 pristine high-provenance videos, approximately 100,000 manipulated images, and 4,000 manipulated videos.
The manipulated-image component draws from approximately 5,000 image-manipulation journals, according to MFC Datasets: Large-Scale Benchmark Datasets for Media Forensic Challenge Evaluation. The 2018 period is essential context: these are legacy benchmark-description figures, not a claim that the collection has remained unchanged.
The DARPA MediFor releases show how storage can differ even when image counts are relatively modest. The NC16 kickoff image dataset contains 1,200 images and is about 4 GB. The NC17 evaluation image dataset contains about 4,000 images and is about 41 GB, while the NC17 evaluation video dataset contains 360 videos and is about 117 GB. These release figures are listed by OpenMFC.
| Media-forensics collection | Content | Reported scale | Period |
|---|---|---|---|
| NIST MFC world-data component | Internet-sourced images | 35 million | 2018 description |
| NIST MFC world-data component | Video clips | 300,000 | 2018 description |
| DARPA MediFor NC16 | Images; storage | 1,200; about 4 GB | NC16 release |
| DARPA MediFor NC17 | Images; storage | About 4,000; about 41 GB | NC17 release |
| DARPA MediFor NC17 | Videos; storage | 360; about 117 GB | NC17 release |
The comparison reinforces that media count is not a substitute for storage volume. Resolution, format, metadata, and video duration can change processing requirements substantially.
Remote-sensing archives
Remote-sensing image processing is driven by continuous acquisition and archive infrastructure. Landsat Project Statistics, updated June 2026, reports about 1 TB of new Landsat mission data acquired and downlinked each day. Landsat processing adds about 3 TB to the USGS EROS archive daily.
The same USGS page reports nearly 250,000 active user accounts accessing Landsat products over the past decade. EROS Legacy reports that the EROS archive passed 100 million downloaded Landsat scenes and stores more than 74 petabytes of digital data. The daily inflow, cumulative downloads, and current storage are distinct operational measures.
For a dated historical reference, the study The global Landsat archive: Status, consolidation, and direction reported that the archive held 5,532,454 images and 4,134 TB of data on January 1, 2015. Its sensor-specific counts were 1,299,626 MSS images, 1,988,982 Thematic Mapper images, 1,858,501 ETM+ images, and 385,345 OLI images. Those figures retain their January 1, 2015 measurement date and are not presented as current totals.
Medical image-processing evaluation
Model evaluation adds another layer to image-processing statistics: training volume, associated clinical labels, external test volume, and accuracy answer different questions. In a 2026 study described by Machine learning analysis of CT scans, the NIH-described Merlin model was trained on more than 15,000 abdominal CT scans and nearly 1 million associated diagnosis codes.
Merlin was tested on more than 50,000 abdominal CT scans from other hospitals and public datasets. It predicted the correct diagnosis with more than 81% average accuracy across hundreds of diagnosis codes and exceeded 90% accuracy for 102 diagnosis codes. The external test scans and the training scans are different quantities, while the accuracy values describe performance rather than dataset size.
The National Library of Medicine’s Visible Human Project reports that about 4,000 licensees from 66 countries were authorized before access became license-free in 2019. This is a project-history measure of international access, not a scan count or model-performance statistic.
Across these examples, the useful unit depends on the question. Image counts describe collection scale; annotation counts describe supervision; cases and series describe biomedical organization; terabytes and petabytes describe infrastructure; and accuracy describes an evaluated model. The periods range from 2015 and 2018 legacy releases to 2026 repository and operational snapshots, so each figure should be interpreted within its stated source and date.