Galar Dataset for VCE Machine Learning
- Galar is a public video capsule endoscopy dataset comprising over 3.5M labeled images from 80 studies with detailed anatomical and pathological annotations.
- It supports diverse tasks such as binary anomaly detection, multi-label temporal classification, and joint localization through various task-specific label projections.
- The dataset’s rich supervision enables advanced system-level studies, including compression, visibility analysis, and energy-aware control in VCE applications.
Searching arXiv for the original Galar dataset paper and papers that use it.
Searching Galar video capsule endoscopy dataset
Galar is a public video capsule endoscopy (VCE) dataset for machine learning on gastrointestinal imaging. In downstream literature, it is described as a large multi-label VCE resource with 3,513,539 labeled images from 80 patient VCE studies, containing annotated anatomical and pathological findings, and supporting tasks ranging from binary anomaly detection to joint gastrointestinal localization, temporal multi-label classification, and study-level systems evaluation (Werner et al., 8 Apr 2025, Werner et al., 31 Jul 2025, Bause et al., 28 Apr 2026).
1. Scope, modality, and dataset role
Galar belongs to the VCE domain, in which a swallowable capsule camera acquires long image sequences through the gastrointestinal tract. Later task papers treat it not as a small curated image collection but as a dataset of whole studies, which makes it suitable for both frame-level classification and temporally structured inference. One study describes it as “the only publicly available VCE dataset” with labels for both different GI organs and anomalies, while another describes it as the currently largest available VCE dataset and emphasizes that it contains 80 complete studies (Werner et al., 31 Jul 2025, Bause et al., 28 Apr 2026).
The dataset is used across several distinct problem formulations. In anomaly-detection work, Galar is treated as a source of small-intestine capsule frames with pathology annotations. In multi-task work, it becomes a joint benchmark for anatomical localization and binary anomaly detection. In competition-oriented temporal classification, it is the basis for 17-label multi-label temporal prediction with event-level evaluation. This breadth of usage indicates that Galar is not merely an image archive; it functions as a general-purpose supervised benchmark for VCE learning pipelines (Werner et al., 8 Apr 2025, Chary et al., 18 Mar 2026, Won et al., 21 May 2026).
The source images are also reported under multiple operational resolutions. One anomaly-detection study states that Galar videos were originally recorded with varying resolutions from to pixels, with training images resized to and test-time evaluation performed on original image sizes (Werner et al., 8 Apr 2025).
2. Annotation structure and task-specific label projections
Published uses of Galar make clear that the dataset is fundamentally multi-label, but that downstream works frequently project it into task-specific label spaces. In the competition-oriented setting, the frame-level label space comprises 17 classes: 8 anatomical regions and 9 pathological findings. The anatomy classes are mouth, esophagus, z-line, stomach, pylorus, small intestine, ileocecal valve, colon; the pathology classes are active bleeding, angiectasia, blood, erosion, erythema, hematin, lymphangioectasis, polyp, ulcer (Chary et al., 18 Mar 2026, Won et al., 21 May 2026).
Other works collapse this richer structure. For pooled anomaly detection in the small intestine, one paper relabels Galar into a binary problem with Normal Clean Mucosa as the negative class and Polyp, Blood, Active Bleeding, Angiectasia, Erosion, Erythema, Ulcer merged into the positive class (Werner et al., 8 Apr 2025). A multi-task study uses a different projection: five sequential GI organ classes—Mouth, Esophagus, Stomach, Small Intestine, Colon—plus a binary anomaly label formed by combining all pathologies into a single positive class (Werner et al., 31 Jul 2025).
A separate systems paper identifies an additional annotation layer: a technical label group for six studies, centered on image quality. In that subset, selected frames are labeled as good view, reduced view, or no view, with visibility reduction attributed to bubbles or dirt. The definitions are explicit: good view means visibility reduction of less than 50%; reduced view means between 50% and 95%; no view means over 95% (Bause et al., 28 Apr 2026).
| Task projection | Labels used | Notes |
|---|---|---|
| Multi-label temporal classification | 17 classes: 8 anatomy + 9 pathology | Frame-level prediction with temporal event generation |
| Binary anomaly detection | Normal Clean Mucosa vs pooled pathology class | Small intestine only |
| Multi-task localization + anomaly | 5 organ classes + binary anomaly | All pathologies collapsed into one anomaly class |
| Technical quality analysis | good view / reduced view / no view; bubbles or dirt | Available for 6 studies |
This multiplicity of label projections is a defining property of Galar’s published use. It suggests that the underlying annotations are rich enough to support distinct supervision regimes, but also that results from different papers are not directly comparable unless the relabeling protocol is specified.
3. Splits, partitions, and evaluation protocols
Galar’s published train/validation/test usage is notably protocol-dependent. For pooled anomaly detection, the test set preserves patient IDs $61$–$80$ from the original publication, while patient IDs $1$–$60$ are split 80:20 (train:val) on a patient-wise basis (Werner et al., 8 Apr 2025). This protocol is intended to avoid leakage across adjacent frames from the same study.
A multi-task study retains the same high-level structure—IDs $61$–$80$ as untouched test studies and IDs $1$–0 as a development pool—but after preprocessing and downsampling it performs a 70:30 training/validation split on the development data. That paper states that the data were “randomly shuffled and split,” but does not specify whether train/validation separation was again patient-wise at that stage (Werner et al., 31 Jul 2025).
Competition-oriented temporal classification papers describe a different partitioning vocabulary. One reports a training partition of 60 videos with approximately 494,696 sampled frames, a validation partition of 10 videos with approximately 80,399 frames, and a held-out RARE-VISION test set comprising 3 NaviCam examinations with 161,025 frames in total (Chary et al., 18 Mar 2026). Another temporal paper reports results on the same three videos—ukdd_navi_00051, ukdd_navi_00068, and ukdd_navi_00076—but does not provide a full dataset inventory beyond that benchmark framing (Won et al., 21 May 2026).
These published protocols indicate that Galar benchmarking is not organized around a single universally adopted split. A plausible implication is that methodological comparisons require careful attention to the exact task projection, patient partitioning rule, and evaluation backend.
4. Benchmark tasks and reported results
Reported Galar benchmarks span binary classification, multi-task frame classification, and temporal event detection. Because the label spaces and metrics differ, these results describe different operating regimes rather than a single leaderboard.
| Task/setup | Reported best result | Source |
|---|---|---|
| Binary anomaly detection over pooled small-intestine classes | AUC 76.98% with SVM-based ensemble | (Werner et al., 8 Apr 2025) |
| Joint 5-organ localization + binary anomaly detection | 93.63% localization accuracy and 87.48% anomaly-detection accuracy, using 1M parameters | (Werner et al., 31 Jul 2025) |
| Multi-label temporal classification on held-out RARE-VISION test set | temporal [email protected] = 0.2456, [email protected] = 0.2353 | (Chary et al., 18 Mar 2026) |
| Post-competition temporal model (“GALAR-TemporalNet v2”) | [email protected] = 0.3409, [email protected] = 0.3333 | (Won et al., 21 May 2026) |
In the pooled anomaly-detection formulation, the strongest ranking metric comes from an SVM ensemble combining at least classifier and autoencoder signals, with AUC 76.98, recall 80.58, accuracy 73.83, F1 27.44, MCC 28.29, and precision 16.53. The plain image classifier is stronger on accuracy, F1, MCC, and precision, with accuracy 87.28, F1 37.01, MCC 34.45, and precision 26.59, while the standalone autoencoder performs poorly on Galar with AUC 45.83 and MCC -4.15 (Werner et al., 8 Apr 2025).
In the multi-task formulation, the shared MobileNetV3-Small model produces large gains for localization when coupled with sequence constraints. The best reported configuration, MT DWA Focal + HMM, reaches 93.63 accuracy, 92.41 F1, 90.40 precision, and 94.94 recall on localization. For anomaly detection, the best reported F1 is 54.38 under MT DWA, while the headline abstract value for anomaly-detection accuracy is 87.48% (Werner et al., 31 Jul 2025).
Competition papers emphasize that rare pathological classes remain difficult even with strong backbones and temporal post-processing. One validation table reports high AP for common anatomical classes such as stomach (0.849), small intestine (0.969), and colon (0.996), but near-zero AP for ultra-rare findings such as active bleeding (0.002), erythema (0.001), hematin (0.001), and ulcer (0.012) (Chary et al., 18 Mar 2026). Another anomaly-detection study similarly notes that blood, active bleeding, and angiectasias are detected relatively well, whereas polyps and erosions are mostly not correctly classified (Werner et al., 8 Apr 2025).
5. Temporal structure, visibility labels, and systems-level studies
A distinctive line of work uses Galar not only for recognition but also for compression, visibility estimation, and energy-aware control in capsule systems. In that setting, the dataset’s full-study organization and technical annotations become central.
One systems paper restricts Galar to the small intestine, then applies offline post-processing to estimate bubble statistics because the dataset does not contain labels for the number, positions, or dimensions of bubbles. On this small-intestine-constrained subset, it reports 1,375,918 frames, of which 869,147 (63.17%) contain bubbles, with 15,742,534 detected bubbles in total. The reported mean bubble coverage is 14.86%, the median is 7.21%, the mean bubble radius is 24.15 px, and the median radius is 23 px (Bause et al., 28 Apr 2026).
The same study uses Galar to validate a compression-based proxy for diagnostic visibility. It reports a compression result of CR = 5.748—equivalent to 82.6% compression—at PSNR = 40.30 dB, and argues that compression ratio is inversely related to bubble coverage. In its operational interpretation, frames with CR < 4 are predominantly obscured by bubbles and medically irrelevant (Bause et al., 28 Apr 2026).
Because Galar contains 80 complete studies, it also supports study-level simulation. Using the small intestine section of all 80 studies and assuming a constant baseline of 2 fps, the same paper reports a mean whole-system energy reduction of 20.58% from compression alone, and an average reduction of 23.12% when compression is combined with bubble-aware frame-rate adaptation. The greatest observed reduction is 40%, reported for study 59 at CR Threshold = 3.6 and reduced fps = 0.67. The conclusion restates this as an additional reduction of 10.99%, without missing any pathologies (Bause et al., 28 Apr 2026).
These results show that Galar’s value extends beyond recognition benchmarks. Its longitudinal structure and technical labels permit system-level studies in which temporal continuity, visibility degradation, and resource constraints are evaluated jointly.
6. Class imbalance, reproducibility issues, and interpretive caveats
A recurring feature across all published uses is extreme class imbalance. One competition paper states that pathological findings constitute less than 0.1% of all annotated frames, and reports rare-class supports in the training partition as low as 3 positive frames for z-line, 73 for mouth, and 150 for active bleeding, with positive-to-negative ratios exceeding 1:3,000 for some labels (Chary et al., 18 Mar 2026). A later temporal paper likewise emphasizes extreme imbalance, long-range temporal dependencies, and pathology–anatomy entanglement as the three core challenges of the benchmark (Won et al., 21 May 2026).
Several methodological caveats also recur. The anomaly-detection paper notes that Galar “only comprises labeled frames,” so unlabeled-data advantages available in other VCE datasets cannot be exploited; this is used to explain why the autoencoder branch contributes less on Galar than on Kvasir-Capsule (Werner et al., 8 Apr 2025). The compression study points out that Galar provides only post-processed, demosaicked RGB images, not native RAW Bayer sensor data, so Bayer inputs must be reconstructed by RGGB subsampling before compression experiments (Bause et al., 28 Apr 2026).
Dataset characterization is also fragmented across the literature. Some papers emphasize the 80 patient studies and the 3,513,539 labeled images (Werner et al., 8 Apr 2025), others foreground challenge partitions such as 60 training videos, 10 validation videos, and a 3-exam held-out test set (Chary et al., 18 Mar 2026), and method-centric temporal papers do not supply a full standalone dataset card (Won et al., 21 May 2026). This suggests that reproducing published Galar results generally requires reading task-specific benchmark papers rather than relying on a single harmonized specification.
Taken together, these caveats position Galar as a powerful but nontrivial benchmark. It is unusually rich in supervision for VCE, especially when anatomy, pathology, temporal continuity, and image quality are considered together, but its effective use depends on explicit relabeling choices, careful split management, and methods that can tolerate severe long-tail imbalance and strong temporal correlation.