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.

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

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

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. VQA result summary. Depth ordering accuracies of VLMs are close to random guess and inferior to DepthAnythingV2 baseline. VLMs’ language consistency has a wide spread.

Uses

To load an image subset:

from datasets import load_dataset

# Load a specific subset
# (For all available subsets, see the Dataset Structure section below)
image_dataset = load_dataset("liuyiqian/O3-D", "kb-1cue", split="train")

# Access the first sample
sample = image_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

To load visual questions:

from datasets import load_dataset

vqa_dataset = load_dataset("liuyiqian/O3-D", "visual_questions", split="train")

# Access the first sample
sample = vqa_dataset[0]

subset_name = sample['subset_name']
image_name = sample['image_name']
is_marked = sample['is_marked']

question = sample['question']

To retrieve visual questions for a specific image:

from datasets import load_dataset

image_subset_name = "kb-1cue"
image_dataset = load_dataset("liuyiqian/O3-D", image_subset_name, split="train")

vqa_subset_name = "visual_questions"
vqa_dataset = load_dataset("liuyiqian/O3-D", vqa_subset_name, split="train")

image_sample = image_dataset[0]
image_name = image_sample['image_name']

# set this to True when using the `image_sample['marked_image']` image field; otherwise False
is_marked = False

# As one lookup approach, use pd.DataFrame's index 
vqa_df = (
    vqa_dataset.to_pandas()
        .set_index(["subset_name", "image_name", "is_marked"])
        .sort_index()
)

retrieval_key = (image_subset_name, image_sample['image_name'], is_marked)
filtered_vqa_df = vqa_df.loc[retrieval_key]

Dataset Structure

Image subsets

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 fields:

• image-based fields
  • image / augmented_image: main image field, i.e. an RGB image of O3-D scene
  • marked_image: same image as above, but with markers
    • to save space, marked images of the kb_* subsets are not available on HF; to generate them, see https://github.com/lyiqian/o3-d
  • depth_map: metric depth map in meters (for kb_* subsets only)
  • targ_seg: segmentation mask of the target
  • dist_seg: segmentation mask of the distractors
• important tabular data fields
  • image_name, image file name
  • odd_position: the position of the odd/target object, either "near", "far", or "none" (only in kb-0cue subset).
    • serve as ground truth for depth ordering VQA
  • cues: a list of controlled cues (see Glossary)
    • for the mixed cue real-mcue subset, the 2 most prominent cues
  • cue_strength, regular or double cue strengths
• other tabular fields
  • depth_scale: min and max depth values (in meters)
  • env_cat, environment category
  • env_id, environment ID
  • obj_cat, object category
  • obj_id, object ID
• real-mcue subset fields
  • we included original dataset fields from O3; for their descriptions, see O3
  • {targ,dist}_{min,med,max}_depth, estimated min/median/max depth using targ_seg/dist_seg masks via zoedepth

Visual question subset

The corresponding visual questions for the image subsets described above, available as subset visual_questions. The subset_name, image_name, and is_marked fields combined serve as a retrieval key given an image in an image subset.

Dataset fields:

• image_name
• subset_name
• is_marked
• question
• ques_clarity
• icl, in-context learning prompt
• cot, chain-of-thoughts prompt

Dataset Creation

Also see methodology section in GitHub: https://github.com/lyiqian/o3-d.

Curation Rationale

We create O3-D to study how well vision-language models can

• understand referring expressions; and
• perceive depth ordering by utilizing pictorial depth cues.

Source Data

Main image source: rendered images with Kubric

Other image source: real-world images

• captured by DSLR camera
• selected from O3

Visual question source: sampled from a template-based prompt formatting process

Data Collection and Processing

Selection criteria for O3 images. To ensure that the images are suitable for equivalent depth ordering questions, we select 171 images from O3 where the odd target was behind or in front of all the distractors.

Annotations of O3 images. As real-world images in O3 have mixed pictorial cues, we label them by providing two most prominent cues.

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

BitTex

@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},
}