Pushing the limits of Visual Grounding: Pre-training on large synthetic datasets

Visual Grounding is a crucial computer vision task requiring a deep understanding of data semantics. Leveraging the transformative trend of training controllable generative models, the research aims to demonstrate the substantial improvement of state-of-the-art visual grounding models through the use of massive, synthetically generated data. The study crafts a synthetic dataset using controllable generative models, offering a scalable solution to overcome challenges in traditional data collection processes. The study introduces a synthetic dataset, employing controllable generative models for scalability. Evaluating visual grounding model (TransVG) — on the synthetic dataset showcases promising results, with attributes contributing to a diverse dataset of 250,000 samples. The resulting datasets showcases the impact of synthetic data on visual grounding evolution, contributing to advancements in this dynamic field.