microsoft/Do-You-See-Me

DoYouSeeMe

Overview

The DoYouSeeMe benchmark is a comprehensive evaluation framework designed to assess visual perception capabilities in Machine Learning Language Models (MLLMs). This fully automated test suite dynamically generates both visual stimuli and perception-focused questions (VPQA) with incremental difficulty levels, enabling a graded evaluation of MLLM performance across multiple perceptual dimensions. Our benchmark consists of both 2D and 3D photorealistic evaluations of MLLMs.

Theoretical Foundation

The dataset's structure is grounded in established human psychological frameworks that categorize visual perception into core abilities (Chalfant and Scheffelin, 1969). Drawing inspiration from standardized assessments like the Test of Visual Perception Skills (TVPS) (Gardner, 1988) and Motor-Free Visual Perception Test (MVPT) (Colarusso, 2003), DoYouSeeMe adapts these principles to create a systematic evaluation methodology for machine vision systems.

Perceptual Dimensions

The benchmark focuses on seven key dimensions of visual perception:

1. Shape Discrimination (2D and 3D): Evaluates the ability to recognize shapes.

2. Joint Shape-Color Discrimination (2D and 3D): Evaluates the ability to jointly recognize shapes and color.

3. Visual Form Constancy (2D and 3D): Tests MLLM ability to identify a test shape configuration from similarly placed disctractors.

4. Letter Disambiguation (2D and 3D): Tests the recognition of letters.

5. Visual Figure-Ground (2D): Evaluates the ability to distinguish the main object from its background under varying conditions.

6. Visual Closure (2D): Assesses the ability to complete partially obscured shapes by mentally filling in missing information.

7. Visual Spatial (2D and 3D): Examines the ability to perceive positions of objects relative to oneself and to other objects.

Note: While human visual perception also includes Visual Memory (the ability to remember sequences of presented images), this dimension is omitted from the benchmark as current MLLMs lack short-term visual memory capabilities beyond textual descriptions.

Technical Implementation

The entire dataset generation framework is implemented in Python and uses SVG representations to create visual stimuli with precisely controlled parameters. This approach allows for:

• Dynamic generation of test images with systematic variations
• Controlled difficulty progression across perception dimensions
• Reproducible evaluation conditions
• Fine-grained assessment of model performance

Control Parameters

The code is open-sourced to facilitate further research and advancement in the field of visual perception for artificial intelligence systems.

Paper: DoYouSeeMe Benchmark on arXiv

Code: DoYouSeeMe Github Repo

Samples

Visual Spatial

Tests the ability to perceive and understand spatial relationships between objects. Evaluates orientation discrimination and positional awareness.

Sample Question: Starting from the black circle at position (row 1, column 3), how many triangles are there bottom of it in the same row?

Visual Figure-Ground

Examines the ability to distinguish an object from its background. Challenges perception by varying contrast, noise, and complexity.

Sample Question: The figure consists of a Target image, which is embedded in some background noise. Out of the four given options, your task is to pick the option which has the same figure as the target image. Respond as follows: Option <your answer (choose between 1, 2, 3, or 4)>.

Visual Form Constancy

Assesses recognition of shapes despite changes in size, orientation, or context. Tests invariance in visual perception.

Sample Question: The figure consists of a Target image. Out of the four given options, your task is to pick the option which has the same figure as the target image. Respond as follows: Option <your answer (choose between 1, 2, 3, or 4)>.

Shape Disambiguation

Challenges the ability to identify ambiguous shapes that can be interpreted in multiple ways. Explores perceptual flexibility.

Sample Question: Count the total number of triangles in the image, including each concentric triangle separately. For example, if there is one triangle with 2 inner concentric rings, that counts as 3 triangles. Respond with only a number.

Shape Color Discrimination

Tests the ability to differentiate shapes based on color properties while controlling for other visual features.

Sample Question: Count the number of star's that are red.

Letter Disambiguation

Examines recognition of letters under various transformations and distortions. Evaluates robustness of character recognition.

Sample Question: The image shows one or more letters formed by a grid of small squares. What letter(s) can you identify in this image? Please respond with only the letter(s) you see.

Visual Closure

Tests the ability to recognize incomplete figures by mentally filling in missing information. Evaluates gestalt processing.

Sample Question: The figure consists of a target image which is complete, Out of the four given options (which are partially complete), your task is to pick the option which when completed matches the target image. Respond as follows: Option <your answer (choose between 1, 2, 3, or 4)>.

Citation

If you use this benchmark or dataset in your research, please cite our work as follows:

@misc{kanade2025multidimensionalbenchmarkevaluating,
      title={Do You See Me : A Multidimensional Benchmark for Evaluating Visual Perception in Multimodal LLMs}, 
      author={Aditya Kanade and Tanuja Ganu},
      year={2025},
      eprint={2506.02022},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.02022}, 
} 

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

License 📜

The code in this repository is licensed under the MIT License. The dataset is licensed under the Community Data License Agreement - Permissive - Version 2.0 (CDLA-Permissive-2.0).