AA-SGAN: Adversarially Augmented Social GAN with Synthetic Data

TL;DR

AA-SGAN addresses pedestrian trajectory forecasting by leveraging synthetic data through a learnable synth-augmentation module trained in an end-to-end adversarial loop. The architecture couples an Augmenter, a Generator, and a Discriminator to produce synth-augmented trajectories that diversify training while preserving realism, enabling improved predictions on real-world data. Experimental results on ETH/UCY show substantial gains over baselines trained on real, synthetic, or hybrid data, with ADE and FDE reductions of about 20% and 25%, respectively. Ablation studies demonstrate that joint end-to-end training and maintaining a balanced real-to-synthetic ratio are essential for achieving these gains. The approach suggests that high-quality synthetic data, when properly augmented, can meaningfully boost pedestrian path prediction in real-world scenarios.

Abstract

Accurately predicting pedestrian trajectories is crucial in applications such as autonomous driving or service robotics, to name a few. Deep generative models achieve top performance in this task, assuming enough labelled trajectories are available for training. To this end, large amounts of synthetically generated, labelled trajectories exist (e.g., generated by video games). However, such trajectories are not meant to represent pedestrian motion realistically and are ineffective at training a predictive model. We propose a method and an architecture to augment synthetic trajectories at training time and with an adversarial approach. We show that trajectory augmentation at training time unleashes significant gains when a state-of-the-art generative model is evaluated over real-world trajectories.

Figures (5)

• Figure 1: Examples of real (a), synthetic (b), and synth-augmented (c) trajectories by the AA-SGAN Augmenter.
• Figure 2: Social GAN for pedestrian trajectories prediction gupta2018social adversarial architecture. The Generator learns to generate real-looking trajectories, the Discriminator learns to tell real from generated trajectories. A social pooling module and an ad-hoc loss function enforce the social plausibility of generated trajectories.
• Figure 3: AA-SGAN for adversarially augmented pedestrian trajectories prediction architecture. The Augmenter learns to augment synthetic trajectories into synth-augmented; the Generator learns to generate trajectories prediction; the Discriminator learns to discriminate real from generated and synth-augmented trajectories.
• Figure 4: Examples of frames from the JTA dataset (top) and the corresponding trajectories (bottom) we used as a synthetic training set.
• Figure 5: Visual comparison between the groud-truth (blue) and predicted trajectory by SGAN with real (green), synthetic (magenta) and hybrid (yellow) training. Predictions by the proposed AA-SGAN are in red and are the closest to the ground truth. The selected results are taken from univ and zara1 testset.