Synset Signset Germany: a Synthetic Dataset for German Traffic Sign Recognition

TL;DR

The paper introduces Synset Signset Germany, a large synthetic dataset for German traffic sign recognition generated via a hybrid pipeline that combines GAN-based texture synthesis with analytic scene modulation. The 105,500-image dataset spans 211 classes and includes rich per-image metadata, segmentation masks, and a subset aligned to GTSRB. Across in-domain and cross-domain evaluations, models trained on the synthetic data achieve strong performance, with notable robustness and explainability analyses enabled by the dataset's controllable parameters. The work demonstrates the value of scalable, explainable synthetic data for traffic sign recognition while outlining limitations and avenues for extending the approach to international signs and enhanced realism.

Abstract

In this paper, we present a synthesis pipeline and dataset for training / testing data in the task of traffic sign recognition that combines the advantages of data-driven and analytical modeling: GAN-based texture generation enables data-driven dirt and wear artifacts, rendering unique and realistic traffic sign surfaces, while the analytical scene modulation achieves physically correct lighting and allows detailed parameterization. In particular, the latter opens up applications in the context of explainable AI (XAI) and robustness tests due to the possibility of evaluating the sensitivity to parameter changes, which we demonstrate with experiments. Our resulting synthetic traffic sign recognition dataset Synset Signset Germany contains a total of 105500 images of 211 different German traffic sign classes, including newly published (2020) and thus comparatively rare traffic signs. In addition to a mask and a segmentation image, we also provide extensive metadata including the stochastically selected environment and imaging effect parameters for each image. We evaluate the degree of realism of Synset Signset Germany on the real-world German Traffic Sign Recognition Benchmark (GTSRB) and in comparison to CATERED, a state-of-the-art synthetic traffic sign recognition dataset.

Figures (8)

• Figure 1: Example images of Synset Signset Germany including challenging conditions as, e.g., noisy, night, overexposed, or shadowed images (lower row).
• Figure 2: Comparison of real images represented by GTSRB (left) and state-of-the-art image synthetization methods for traffic sign recognition. For achieving a better comparability we cropped the images to a similar area if necessary. The DCGAN, LSGAN, and WGAN samples stemming from SynTrafSignRec_GanApproach result from training 200 epochs respectively.
• Figure 3: Overview of the generation pipeline built in OCTANE. Ideal images (a) are procedurally converted to template images (b) defining color degradation and wear/tear masks. A GAN trained on worn traffic signs converts these to diffuse textures (c) that are combined into a 3D scene for physically-based rendering. Segmentation and mask images (d) are rendered using OGRE, while Cycles is used for geometric raytracing of HDR raw image, albedo, and normal image (e). The latter are used to denoise the raytracing samples. Based on this, imaging artifacts are computed on the 2D image data (f).
• Figure 4: GAN setup for defect synthetization.
• Figure 5: Overview of the signs in Synset Signset Germany. The first row of 43 signs corresponds to the classes in GTSRB. Sign shapes are based on the Wikipedia overview of German traffic signs from 2017 onwards (cf. \ref{['fn:wikipedia']}).