The Multi-View Evaluation Protocol for Glass Container Inspection (MVEP) is a specialized benchmark dataset designed to evaluate multi-view fusion methods for industrial quality control of transparent materials. MVEP addresses the critical challenge of automated defect detection and severity assessment on glass container surfaces, where view-dependent optical phenomena (including specular reflections, refractions, and transparency effects) significantly limit the reliability of single-view inspection. The dataset comprises 16,000 synchronized multi-view images of glass containers captured from six calibrated viewpoints under controlled industrial lighting conditions. Each container is annotated with object-level bounding boxes and ordinal severity labels corresponding to surface degradation defects, ranging from minimal visual alteration to critical damage requiring rejection. The defect taxonomy focuses on erasure-type surface defects, replacing earlier wipe-based definitions to better reflect realistic industrial degradation patterns, and excludes any reference to scuffing. MVEP provides a realistic distribution of quality grades encountered in production environments and includes inherent annotation uncertainty due to the subjective nature of visual quality assessment. The dataset is particularly well suited for evaluating ordinal classification methods, multi-view fusion strategies, cross-view consistency constraints, and robustness to annotation noise in industrial inspection scenarios involving transparent materials.
The dataset is organized using a structured directory layout designed for defect detection tasks. The data are split into two main directories, train and test, each containing samples organized at the article level. Each article is represented by 12 images captured from multiple viewpoints, with a corresponding annotation file provided for each image.
All images follow a standardized naming convention defined as:
timestamp_modality_view_camera.extension.
For example: 20260109091529549_26_10_E.bmp, where the filename encodes the acquisition timestamp, an acquisition identifier, the view index, and the camera orientation. The camera is indicated by a single character, with E denoting a top-down camera and C a front-facing camera. This naming scheme ensures unique identification of each image and enables straightforward association with metadata and annotation files.
Annotations are provided in YOLO-style text format, with each line corresponding to a single defect instance. The annotation format is:
class_id x_center y_center width height,
where class_id represents the defect category and the bounding box coordinates are normalized by the image width and height. For example:
0 0.261557 0.478477 0.058647 0.355315.
In addition, a dataset.json file located at the root directory contains detailed metadata for each image. This file includes the image identifier, file name and path, image resolution, view index, camera type, capture configuration, article position, defect class information, and bounding box annotations expressed in absolute pixel coordinates. This metadata file enables robust traceability, supports multiple annotation formats, and facilitates advanced analysis and dataset integration.
{"image_id": "20260109091529549_26", "image_name": "20260109091529549_26_10_E.bmp", "image_path": "20260109091529549_26_10_E.bmp", "width": 1008, "height": 1008, "view": 10, "camera": "E", "capture_type": "Shoulder", "pattern": 1, "position": "B", "class_id": 0, "class_name": "defect_class_0", "bbox": { "x": 0.0, "y": 305.998046875, "w": 51.50910949707031, "h": 300.519775390625 } },