2018/06/14 by Mohsen Ghafoorian, Ghafoorian, Mohsen, Cedric Nugteren +7 · 1 citation
Computer Science · Engineering · #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #Autonomous Vehicle Technology and Safety #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.1806.05525
openalex publication_date 2018/06/14 · openalex created_date 2022/09/19 · openalex updated_date 2026/07/28
Convolutional neural networks have been successfully applied to semantic\nsegmentation problems. However, there are many problems that are inherently not\npixel-wise classification problems but are nevertheless frequently formulated\nas semantic segmentation. This ill-posed formulation consequently necessitates\nhand-crafted scenario-specific and computationally expensive post-processing\nmethods to convert the per pixel probability maps to final desired outputs.\nGenerative adversarial networks (GANs) can be used to make the semantic\nsegmentation network output to be more realistic or better\nstructure-preserving, decreasing the dependency on potentially complex\npost-processing. In this work, we propose EL-GAN: a GAN framework to mitigate\nthe discussed problem using an embedding loss. With EL-GAN, we discriminate\nbased on learned embeddings of both the labels and the prediction at the same\ntime. This results in more stable training due to having better discriminative\ninformation, benefiting from seeing both `fake' and `real' predictions at the\nsame time. This substantially stabilizes the adversarial training process. We\nuse the TuSimple lane marking challenge to demonstrate that with our proposed\nframework it is viable to overcome the inherent anomalies of posing it as a\nsemantic segmentation problem. Not only is the output considerably more similar\nto the labels when compared to conventional methods, the subsequent\npost-processing is also simpler and crosses the competitive 96% accuracy\nthreshold.\n