Aircraft Images with Hierarchical Model Labels and Bounding Boxes

Containing 10,200 images of aircraft, with 100 images for each of 102 different aircraft model variants. Each image includes a tight bounding box annotation and a hierarchical label spanning model, variant, family, and manufacturer levels. The data is pre-split into equally-sized training, validation, and test subsets.

Use Cases

• Train an object detection model to localize aircraft using the provided bounding box annotations.
• Perform fine-grained image classification across 102 visually distinct aircraft variant classes.
• Analyze the hierarchical label structure to study visual similarity across 70 families or 41 manufacturers.
• Benchmark multi-task learning models that jointly predict bounding boxes and hierarchical class labels.

Strengths

• Contains 10,200 annotated images, providing a substantial volume for model training.
• Offers a structured four-level hierarchy (102 variants, 70 families, 41 manufacturers) for detailed classification tasks.
• Includes precise bounding box annotations for the main aircraft in each image.
• Provides a pre-defined, equally-sized split for training, validation, and final evaluation.

Limitations

• The dataset is focused primarily on airplanes, which may limit generalization to other aircraft types.
• The 'Model' level of the hierarchy is noted as not used for evaluation due to visual indistinguishability, which may restrict its utility.
• No information is provided on image resolution, lighting conditions, or background diversity, which could affect model robustness.

Provenance

Source

HuggingFaceM4 via Hugging Face platform.

Collection Method

Images annotated with bounding boxes and hierarchical aircraft model labels; specific collection method not detailed.

Freshness

Last updated in June 2022.