ZSAD Shipping Container Dataset

This dataset is specifically curated for fine-tuning Zero-Shot and Prompt-Based Anomaly Detection models (such as Bayes-PFL or AA-CLIP). It moves beyond simple "whole-box" classification by providing granular views of critical shipping container components, allowing models to learn the hierarchical relationship between a global object and its local features. Dataset Overview The dataset consists of high-resolution industrial imagery designed to help models distinguish between "Normal" operational states and "Anomalous" conditions (e.g., dents, structural corrosion, or door seal failures). While many datasets treat a container as a single entity, this collection breaks the asset down into three distinct semantic categories: Container: Global views of the entire unit. Container Door: Focused views of the locking mechanisms and panels. Hinges: Macro-level views of pivot points and attachment hardware. Folder Structure & Data Splits The dataset is organized to support a robust training pipeline where the model learns the general context of the container first, then refines its understanding of specific high-wear components. Folder Name Split Purpose container/ Train / Test / Valid The primary set for evaluating model performance. The train/test/valid sets contain both normal and anomalous samples. container_door/ Train Only Provides localized "normal" state prototypes for door geometry and locking bars. hinges/ Train Only Provides fine-grained visual features of hardware to reduce false positives in high-texture areas. Application in Zero-Shot Fine-Tuning This specific structure is optimized for models like AA-CLIP or Bayes-PFL because: Prompt-Based Learning: The container_door and hinges training folders allow you to generate "normalcy prompts" or prototypes. When the model evaluates a test image in the container folder, it can reference these learned component features to spot localized anomalies. Hierarchical Context: By training on sub-components without a test split, you are essentially providing the model with a "dictionary" of healthy parts. This prevents the model from flagging a complex hinge or a latch as an anomaly simply because it has a high-frequency visual pattern. Anomaly Localization: The split encourages the model to learn that an anomaly on a "hinge" is a subset of an anomaly on a "container," improving the precision of heatmaps or bounding box outputs.