Learning an Optimisable Semantic Segmentation Map with Image Conditioned Variational Autoencoder

Pengcheng Zhuang, Yusuke Sekikawa, Kosuke Hara, Hideo Saito

Abstract: Recent semantic segmentation systems have achieved significant improvement by performing pixel-wise training with hierarchical features using deep convolutional neural network models. While the learning process usually requires pixel-level annotated images, it is difficult to get desirable amounts of fine-labeled data and thus the training set size is more likely to be limited, often in thousands. This means that top methods for a dataset can be fine-tuned for a specific situation, making the generalization ability unclear. In real-world applications like self-driving systems, ambiguous region or lack of context information can cause errors in the predicted results. Resolving such ambiguities is crucial for subsequent operations to be performed safely. We are inspired by work from CodeSLAM where optimizable pixel-wise depth representation is learned. We modify the regression method to work on the pixel-wise classification problem. By training a variational auto-encoder network conditioned with a color image, the computed latent space works as a low-dimensional representation of semantic segmentation, which can be efficiently optimized. As a consequence, our model can correct the error or ambiguity of the prediction during the inference phase given useful scene information. We show how this approach works by giving partial scene truth and perform optimization on the latent variable.