Stereo4D Benchmark for 3D Object Detection with Ground-Truth Depth

A 3D object detection benchmark provides 383 validation images with 2,782 annotations, derived from Stereo4D video data. Created by AllenAI, the dataset offers ground-truth stereo depth for model evaluation. It was last updated in April 2026.

Use Cases

• Benchmark 3D bounding box detection accuracy using the provided 2,782 object annotations.
• Train models to predict depth from stereo image pairs using the ground-truth stereo depth data.
• Evaluate model robustness on real-world driving scenes extracted from Stereo4D videos.
• Fine-tune object detectors on the 383 validation images before hidden test set evaluation.

Strengths

• 2,782 detailed object annotations across the validation set
• 383 images sourced from real-world Stereo4D driving videos
• Includes ground-truth stereo depth for precise 3D localization

Limitations

• Small validation set of only 383 images
• Test set is held out and not available for public analysis, limiting reproducibility
• Represents only 7% of all Stereo4D videos, potentially reducing scene diversity

Provenance

Source

AllenAI, derived from the Stereo4D video dataset.

Collection Method

Images extracted as single frames from Stereo4D test set videos.

Freshness

Last updated April 2026.