CropVLM: Learning to Zoom for Fine-Grained Vision-Language Perception
This addresses fine-grained perception challenges in VLMs for applications like scene-text recognition, offering a low-cost, non-invasive solution without catastrophic forgetting.
The paper tackles the problem of fine-grained image understanding in Vision-Language Models (VLMs) by introducing CropVLM, a method that enables dynamic zooming on relevant image regions, resulting in significant performance improvements on high-resolution tasks, especially for out-of-domain benchmarks.
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.