Slow Perception: Let's Perceive Geometric Figures Step-by-step

Abstract: Recently, "visual o1" began to enter people's vision, with expectations that this slow-thinking design can solve visual reasoning tasks, especially geometric math problems. However, the reality is that current LVLMs (Large Vision LLMs) can hardly even accurately copy a geometric figure, let alone truly understand the complex inherent logic and spatial relationships within geometric shapes. We believe accurate copying (strong perception) is the first step to visual o1. Accordingly, we introduce the concept of "slow perception" (SP), which guides the model to gradually perceive basic point-line combinations, as our humans, reconstruct complex geometric structures progressively. There are two-fold stages in SP: a) perception decomposition. Perception is not instantaneous. In this stage, complex geometric figures are broken down into basic simple units to unify geometry representation. b) perception flow, which acknowledges that accurately tracing a line is not an easy task. This stage aims to avoid "long visual jumps" in regressing line segments by using a proposed "perceptual ruler" to trace each line stroke-by-stroke. Surprisingly, such a human-like perception manner enjoys an inference time scaling law -- the slower, the better. Researchers strive to speed up the model's perception in the past, but we slow it down again, allowing the model to read the image step-by-step and carefully.

• The paper introduces slow perception, a method that decomposes geometric figures into basic units for improved model reasoning.
• It employs a two-stage approach—perception decomposition and sequential line tracing—to ensure precise visual reconstruction.
• Evaluation reveals a 6% F1-score boost and notable gains even with frozen vision encoders, confirming its practical impact.

An Expert Analysis of "Slow Perception: Let's Perceive Geometric Figures Step-by-step"

The paper "Slow Perception: Let's Perceive Geometric Figures Step-by-step" addresses the challenge of geometric figure parsing in computer vision, proposing a novel approach termed "slow perception" (SP). This method is inspired by slow-thinking paradigms and aims to enhance the visual reasoning capabilities of Large Vision LLMs (LVLMs) through a more deliberative and human-like perceptual process.

Core Contributions

The paper introduces the concept of slow perception as a means to overcome the current limitations of LVLMs in accurately interpreting and replicating geometric figures. The proposed method comprises two main stages:

1. Perception Decomposition: This stage involves breaking down complex geometric figures into simpler units, such as basic point-line combinations. The authors argue that this decomposition leads to a unified representation of diverse geometric shapes, which is essential for effective parsing. By modeling geometric perception in this incremental manner, the approach leverages geometry's intrinsic elemental relations.
2. Perception Flow: To address the challenges of accurately tracing lines in geometric figures, this stage employs a technique analogous to a "perceptual ruler," where each line is traced in segments. This method avoids long visual jumps, ensuring a more precise reconstruction of lines by tracing them stroke-by-stroke, akin to how humans draw with a ruler.

Notably, the authors posit an inference time scaling law where the slower the perceptual process, the better the model's performance. This is counterintuitive to the prevailing trend where speeding up models is often prioritized; here, slowing down allows for a more meticulous and thorough understanding of the visual input.

Performance Evaluation

The efficacy of slow perception is evaluated using synthetic and real-world datasets, achieving an F1-score improvement of 6% over baseline methods on the test set. The experimental results highlight that even with the perceptual ruler set to large values, the method consistently enhances performance. A further scaling of the perceptual ruler to smaller values yielded progressively better results, adhering to the proposed inference time scaling law.

The paper also examines the impact of slow perception on existing LVLMs like GOT-OCR2.0, demonstrating substantial improvements across various metrics, even when the vision encoder part is frozen. Additionally, the research underpins the theoretical claims with practical implementations, showcasing how the approach can be adapted to other LVLMs for further validation.

Implications and Future Directions

The investigation into slow perception offers promising insights for improving visual reasoning in AI systems. From a theoretical perspective, it raises intriguing questions about the balance between processing speed and the depth of perceptual understanding. The human-like process introduced here serves as a foundation for advancing AI's ability to replicate complex perception tasks akin to human cognition.

Future research could explore refining this method with reinforcement learning to dynamically adapt the length of the perceptual ruler based on the complexity of geometric tasks. Additionally, extending the principles of slow perception to more generalized computer vision problems could open new avenues for enhancing the capabilities of AI in understanding and interacting with the visual world.

The paper's findings challenge the assumptions underlying rapid model inference, suggesting that slowing down to perceive more thoroughly might hold the key to significant advancements in visual cognition within AI systems. The focus on geometric figures, serving as abstractions of more complex visual phenomena, provides a structured approach to refining AI's perceptual accuracy and understanding.