Meta-Sim: Learning to Generate Synthetic Datasets

TLDR

Training models to high-end performance requires large labeled datasets, which are expensive to obtain. The study aims to automatically synthesize labeled datasets relevant to a downstream task and, when a small validation set is available, to optimize a meta‑objective. Meta‑Sim parametrizes a neural‑network generator that modifies attributes of scene‑graph samples from probabilistic grammars to produce synthetic images and ground‑truth, minimizing the distribution gap to real data and optimizing downstream task performance. Experiments demonstrate that Meta‑Sim markedly improves synthetic content quality compared to a human‑engineered scene grammar, yielding higher downstream task performance.

Abstract

Training models to high-end performance requires availability of large labeled datasets, which are expensive to get. The goal of our work is to automatically synthesize labeled datasets that are relevant for a downstream task. We propose Meta-Sim, which learns a generative model of synthetic scenes, and obtain images as well as its corresponding ground-truth via a graphics engine. We parametrize our dataset generator with a neural network, which learns to modify attributes of scene graphs obtained from probabilistic scene grammars, so as to minimize the distribution gap between its rendered outputs and target data. If the real dataset comes with a small labeled validation set, we additionally aim to optimize a meta-objective, i.e. downstream task performance. Experiments show that the proposed method can greatly improve content generation quality over a human-engineered probabilistic scene grammar, both qualitatively and quantitatively as measured by performance on a downstream task.