Obtaining labeled data for lane detection is expensive and time consuming. We propose a new paradigm for training models to detect lanes in images using unlabeled data from the target domain. and with unaligned labels (labels without their source images). Our method uses adversarial networks in a novel setting. It can be used in multiple training paradigms such as semi-supervised training and domain adaptation from synthetic data and reduces the need for labeled images from the target domain. Our proposed approach, depicted in Figure 1, is trained with unlabeled target domain images (I), and unpaired ground truth annotations (2). In our application, this approach is based on the task-specific assumption that in the road area lane markings correlate with image gradients. Essentially, we train a feature embedding function (3) and skeleton encoder (4) to generate, from an unlabeled input image (1) a gray-scale "lane image" (5) that satisfies two constraints: (a) it looks like a valid ground truth image of lanes and (b) it holds the information required to reconstruct the original gradients in the image. The first constraint is imposed by a discriminator. (6), which tries to distinguish between the generated lane image (5) and unpaired ground truth lane annotations (2). The second is imposed by a decoder (7) that tries to reconstruct the original image gradients (8) from (5). The entire approach, relies on the assumption, that the natural candidate for the encoded image satisfying both constraints is that of the lanes in the scene. The result of the training is the feature extractor generating useful features (9) for lane detection on the target domain.
展开▼
