CropVLM: Learning to Zoom for Fine-Grained Vision-Language Perception

TL;DR

CropVLM is a reinforcement learning-based method that dynamically zooms into relevant image regions to enhance fine-grained vision-language perception without additional supervision.

Contribution

It introduces a low-cost, training-once approach that improves existing VLMs' performance on detailed image understanding tasks without fine-tuning the models.

Findings

Significant performance improvements on fine-grained vision-language tasks.

Effective for out-of-domain benchmarks without modifying the original VLMs.

Does not require human-labeled bounding boxes or synthetic data.

Abstract

Vision-Language Models (VLMs) often struggle with tasks that require fine-grained image understanding, such as scene-text recognition or document analysis, due to perception limitations and visual fragmentation. To address these challenges, we introduce CropVLM as an external low-cost method for boosting performance, enabling VLMs to dynamically ''zoom in'' on relevant image regions, enhancing their ability to capture fine details. CropVLM is trained using reinforcement learning, without using human-labeled bounding boxes as a supervision signal, and without expensive synthetic evaluations. The model is trained once and can be paired with both open-source and proprietary VLMs to improve their performance. Our approach delivers significant improvements on tasks that require high-resolution image understanding, notably for benchmarks that are out-of-domain for the target VLM, without modifying or fine-tuning the VLM, thus avoiding catastrophic forgetting.