Dataset Card for Odd-One-Out Depth (O3-D)

A multi-modal dataset with controlled monocular/pictorial depth cues. O3-D combines odd-out-out and depth-ordering tasks to specifically analyze VLMs' referring expression understanding and basic depth perception. Wherever possible, depth maps and segmentation masks are provided for general usage, such as monocular depth estimation and salient object detection.

Paper: https://arxiv.org/abs/2607.01503 (accepted ECCV2026)

GitHub: https://github.com/lyiqian/o3-d

Dataset Details

O3-D contains

• 4 synthetic and 3 real-world images subsets
• over 1000 unique visual questions with different level of referring clarity
• depth maps for all synthetic images
• segmentation masks for target and distractors

Dataset Description

Fig. 1. O3-D probes VLM depth and language understanding. Each 3D scene contains 5 objects of the same class, one of which (the target) is of different size and placed at a different depth plane. We then generate a number of 2D views with one or two depth cues by controlling the camera, light position, etc. For each image, we pair it with one of the depth-ordering prompt templates, within which we vary the target and distractor descriptions.

Fig. 2. Summarized VQA results. Depth ordering accuracies of VLMs are close to random guess and inferior to DepthAnythingV2 baseline. VLMs’ language consistency has a wide spread.

Uses

from datasets import load_dataset

# Load a specific subset (see Dataset Structure below)
dataset = load_dataset("liuyiqian/O3-D", "kb-1cue", split="train")

# Access the first sample
sample = dataset[0]
rgb_image = sample['image']
depth_map = sample['depth_map']  # Loads as a PIL Image
target_mask = sample['targ_seg']  # Loads as a PIL Image
distractor_segmts = sample['dist_seg']  # Loads as a PIL Image

Dataset Structure

4 synthetic and 3 real-world image subsets:

• kb-0cue, zero cue (negative) baseline
• kb-1cue, images with a single controlled cue
• kb-2cue, images with two controlled cues
• kb-no-lp, a small subset for testing Linear Perspective (LP) cue
• real-012cue, real-world subset with controlled cues for verification purpose
• real-012cue-cropped, cropped subset specifically for depth-ordering task
• real-mcue, real-world subset with uncontrolled cues

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Data Collection and Processing

[More Information Needed]

Annotations [optional]

Annotation process

[More Information Needed]

Bias, Risks, and Limitations

[More Information Needed]

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.

Citation and License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license. You are free to share and adapt the material for any purpose, even commercially, as long as you provide appropriate credit to the original authors.

If you use this dataset in your research, please cite the following paper.

BibTeX:

@misc{liu2026disentanglingpictorialcueunderstanding,
      title={Disentangling Pictorial Cue Understanding from Language Bias in VLMs via Depth Ordering Task},
      author={Yiqian Liu and Iuliia Kotseruba and John K. Tsotsos},
      year={2026},
      eprint={2607.01503},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.01503},
}

Glossary

Controlled pictorial cues:

• OC: Occlusion
• LS: Light and Shadow
• TG: Texture Gradient
• LP: Linear Perspective
• HP: Height-in-Plane
• RS: Relative Size
• FS: Familiar Size
• SA: Saturation
• FO: Focusness