MonoVLM: Monocular 3D Visual Grounding with Vision Language Models

Figure 1. We propose MonoVLM, a simple yet effective method to equip Vision-Language Models (VLMs) with robust monocular 3D grounding capabilities. (a) The model takes an image and the textual query to predict the 3D bounding box (GT and prediction). (b) Even the latest large-scale VLMs struggle to understand 3D structure from 2D images. Our resulting model not only achieves significantly better results than these VLMs but also surpasses specialized vision-only models designed for this task.