VLM's Eye Examination: Instruct and Inspect Visual Competency of Vision Language Models

TL;DR

The paper introduces an eye examination protocol to interrogate how Vision-Language Models perceive visuals, structured around instruction, readiness checks, and examination with the LENS dataset covering color, shape, and semantics. It defines quantitative metrics, SAC and SAS, to evaluate color and shape sensitivity, and uses semantic patch analysis to assess higher-level understanding across two VLMs (LLaVA and InstructBLIP) with varying LLM capacities. Key findings include a consistent insensitivity to green colors driven by the visual encoder, and a dependence of shape and semantic accuracy on the size and capabilities of the LLM component. These insights offer guidance for model design and simple preprocessing strategies to improve real-world tasks like chart understanding while highlighting the need for further exploration of VLM perception and controllability.

Abstract

Vision language models (VLMs) have shown promising reasoning capabilities across various benchmarks; however, our understanding of their visual perception remains limited. In this work, we propose an eye examination process to investigate how a VLM perceives images, specifically focusing on key elements of visual recognition, from primitive color and shape to semantic levels. To this end, we introduce a dataset named LENS to guide a VLM to follow the examination and check its readiness. Once the model is ready, we conduct the examination. Through this examination, we quantify and visualize VLMs' sensitivities to color and shape, and semantic matching. Our findings reveal that VLMs have varying sensitivity to different colors while consistently showing insensitivity to green across different VLMs. Also, we found different shape sensitivity and semantic recognition depending on LLM's capacity despite using the same fixed visual encoder. Our analyses and findings have potential to inspire the design of VLMs and the pre-processing of visual input to VLMs for improving application performance.

Figures (12)

• Figure 1: Eye examination. The process of eye examination contains three steps: instruction, readiness check, and examination. If a VLM has acquainted instructions and is ready, the model conducts examinations of color, shape, and semantic comparisons to assess its visual competency.
• Figure 2: LENS dataset. We visualize the samples in LENS, which is designed to instruct VLM and check its readiness. LENS contains three categories: color, shape, and semantics. Note that questions for each sample are randomly sampled from a pre-defined question set, and the prompts for patches contain two options along with a no answer option. More details and data statistics can be found in the Appendix.
• Figure 3: Visualization Color sensitivity. We measure the sensitivity of VLMs in differentiating between reference and target colors. Visualization of color sensitivity represented by $f(c_{\text{ref}}, c_{\text{target}})$, where $c_{\text{ref}}$ is one of red, green, and blue, and $c_{\text{target}}$ is selected from the color wheel. The result that green has the largest area of the wheel indicates distinguishing green is more challenging compared to red or blue.
• Figure 4: Color correction. The first column is the original images that humans see. The second and third columns show the transformed images based on the color similarity patterns perceived by LLaVA and InstructBLIP, respectively. VLMs see the world greener.
• Figure 5: Shape sensitivity. We measure the sensitivity of VLMs between a circle and target shapes by varying (a) the eccentricity of a circle, (b) the number of vertices in a regular polygon, or (c) the size of a circle. The model is more sensitive if the score changes at (d) lower eccentricity, (e) a higher number of vertices, and (f) a narrower range of size. The results shows that a larger VLM is more sensitive than a smaller one.