WildDet3D Bench: In-the-Wild 3D Object Detection Benchmark

WildDet3D Bench is a 3D object detection benchmark created by AllenAI, last updated in April 2026. It provides a validation split of 2,470 images with 9,256 annotations across 785 categories, aggregated from COCO, LVIS, and Objects365 datasets. The test split is held out for hidden evaluation on a dedicated server.

Use Cases

• Benchmarking 3D object detection model performance based on in-the-wild images from multiple established datasets.
• Training models for multi-category object recognition based on the 785 annotated categories.
• Developing robust detection systems for varied and complex scenes based on the 'in-the-wild' nature of the aggregated data.
• Evaluating model generalization using the held-out test set via the dedicated evaluation server.

Strengths

• Validation set contains 2,470 images with 9,256 annotations, providing a substantial benchmark.
• Aggregates data from three major computer vision datasets (COCO, LVIS, Objects365), likely increasing diversity.
• Covers a wide range of 785 object categories for detection tasks.

Limitations

• Test set is not publicly available, limiting independent full evaluation.
• Column-level documentation and sample data are unavailable, making data structure and semantics unclear prior to download.
• Row count for the full dataset is unknown, which may limit suitability assessment.

Provenance

Source

AllenAI, aggregated from COCO, LVIS, and Objects365 datasets.

Collection Method

Aggregated from existing public computer vision datasets.

Freshness

Last updated 2026-04-07 10:59:22.