MMDOD - Multi-Modal and Multi-View Object Detection Dataset MMDOD - Multi-Modal and Multi-View Object Detection Dataset MMDOD - マルチモーダル・マルチビュー物体検出データセット

The integration of multi-modal and multi-view information has become increasingly important for robust visual inspection, particularly in scenarios involving complex optical media such as glass. While recent advances in image fusion and detection models have shown promising results, progress remains limited by the scarcity of comprehensive datasets that combine multiple imaging modalities and viewpoints with consistent object annotations. In this work, we introduce a novel dataset designed to address this gap by providing multi-modal (Visible, Near-Infrared, Polarization Shift, and Low-Contrast) and multi-view (6 viewpoints per container) image sequences of transparent glass containers under controlled but varied defect scenarios. The dataset features both non-destructive objects (e.g., stickers, filters, elastic bands) and destructive structural anomalies (e.g., scratches), as well as visual artifacts such as mold seams, optical reflections, and shadows, which must be implicitly learned as background. By systematically varying object types, configurations, and orientations, we create a rich benchmark that supports task-driven image fusion for object detection. We also highlight key challenges such as optical distortions due to curvature, object occlusion across views, and modality-specific visibility. This dataset is designed to support the development and evaluation of unified fusion architectures capable of learning from heterogeneous visual inputs, addressing an unmet need in both academic and industrial inspection contexts. The integration of multi-modal and multi-view information has become increasingly important for robust visual inspection, particularly in scenarios involving complex optical media such as glass. While recent advances in image fusion and detection models have shown promising results, progress remains limited by the scarcity of comprehensive datasets that combine multiple imaging modalities and viewpoints with consistent object annotations. In this work, we introduce a novel dataset designed to address this gap by providing multi-modal (Visible, Near-Infrared, Polarization Shift, and Low-Contrast) and multi-view (6 viewpoints per container) image sequences of transparent glass containers under controlled but varied defect scenarios. The dataset features both non-destructive objects (e.g., stickers, filters, elastic bands) and destructive structural anomalies (e.g., scratches), as well as visual artifacts such as mold seams, optical reflections, and shadows, which must be implicitly learned as background. By systematically varying object types, configurations, and orientations, we create a rich benchmark that supports task-driven image fusion for object detection. We also highlight key challenges such as optical distortions due to curvature, object occlusion across views, and modality-specific visibility. This dataset is designed to support the development and evaluation of unified fusion architectures capable of learning from heterogeneous visual inputs, addressing an unmet need in both academic and industrial inspection contexts. マルチモーダルおよびマルチビュー情報の統合は、特にガラスなどの複雑な光学媒体を含むシナリオにおいて、堅牢な視覚検査にとってますます重要になっています。 画像融合と検出モデルの最近の進歩は有望な結果を示していますが、複数の撮像モダリティと視点を一貫したオブジェクトアノテーションと組み合わせた包括的なデータセットの不足により、進歩は依然として制限されています。 この研究では、制御されているが多様な欠陥シナリオの下で、透明なガラス容器のマルチモーダル（可視光、近赤外線、偏光シフト、低コントラスト）およびマルチビュー（容器あたり6つの視点）画像シーケンスを提供することにより、 このギャップに対処するように設計された新しいデータセットを紹介します。データセットには、非破壊的オブジェクト（ステッカー、フィルター、ゴムバンドなど）と破壊的構造異常（傷など）の両方、 および背景として暗黙的に学習する必要がある金型継ぎ目、光学反射、影などの視覚的アーティファクトが含まれています。オブジェクトタイプ、構成、方向を体系的に変化させることにより、 物体検出のためのタスク駆動型画像融合をサポートする豊富なベンチマークを作成します。また、曲率による光学歪み、ビュー間のオブジェクトオクルージョン、モダリティ固有の可視性などの主要な課題も強調しています。 このデータセットは、異種視覚入力から学習できる統一融合アーキテクチャの開発と評価をサポートするように設計されており、学術および産業検査コンテキストの両方で満たされていないニーズに対処します。