Recently, the U-Net-like convolutional neural network architecture composed of encoder, decoder, and symmetric skip-connection has shown impressive performance and has quickly become the de-facto standard in road segmentation tasks. In this letter, a novel Wide-Range Attention Unit (WRAU) is introduced and incorporated into a densely-connected U-Net architecture which we term the Wide-Range Attention U-Net (WRAU-Net). New architecture achieves a wide-range channel-wise attention mechanism by two major contributions. Firstly, we add a partially dense-connections branch to the U-Net architecture to aggregate informative feature map channels with multiple spatial resolutions from different levels of encoder. Secondly, before encoding back into the network, the WRAU is employed at the cluster node of partially dense-connections to learn to weight aggregated feature map channels by a three-layer convolutional network. We evaluate our WRAU-Net on the Massachusetts road segmentation benchmark. The cross-validation results show that the employment of the proposed attention mechanism to classic U-Net architecture gives a tangible boost to test accuracy when compared to U-Net itself and another two state-of-the-art baselines.
Deep models are extremely data hungry. Their success is being driven by the availability of large amounts of data for training. For semantic segmentation tasks on aerial and satellite imagery, a major dilemma at present is that it still relies heavily on manual labelling of data. Among these tasks, the semantic segmentation of road is special since it is possible to use auxiliary data, such as GPS track data, to automatically label data. For a better understanding of this possibility, this paper proposes to rethink some basic issues of labelling approaches for roads. We experimentally investigated the unavoidable class imbalance problem in road segmentation tasks through simulated and real datasets and quantitatively show that class imbalance has a serious detrimental impact on deep model’s generalization performance. We also observed that the detrimental impact even outweighs the benefits of strictly annotating roads – expanding road labels can give deep networks better segmentation accuracy, even though the segmentation location is no longer the edge of the road. We think this is because the impact of class imbalance is much overwhelming than the sensitivity of DNN on the edges of real roads. This finding is valuable for supporting the use of centreline-based approaches in place of edge-based approaches in some applications for better cost-effective solutions. We proposed a guided Random Sample Consensus (RANSAC) algorithm to determine the optimal expansion ratio of road label. On these bases, we further proposed a general framework to combine two networks to achieve better performance than the state-of-the-art performance of using alone. We attribute this to the alleviation of the class imbalance problem because simply cascading the two networks does not achieve the purpose of improving accuracy in our experiments. We believe that this work is enlightening for studies of road segmentation.
Fine-tuning a Deep Convolutional Neural Network (DCNN) model obtained from complicated and tedious pre-training on large-scale datasets seems to have become a standard baseline for various tasks with scale-limited datasets. However, there is little work done to investigate into the pre-training’s impact on fine-tuning. In this letter, we systematically verify that pre-training’s impact originating from large-scale datasets in other domain is still insurmountable for full-training with mainstream datasets in the Remote Sensing Scene Classification (RSSC) domain. However, we confirm that this impact is task-related. That is, once the need of pre-training is satisfied, excessive pre-training may lead to generalization degradation of following fine-tuning. This counter-intuitive finding explains why identical pre-training results in different generalization disparity in various tasks. We also provide experimental evidence to show that excessive or noisy pre-training is possible to make fine-tuning perform worse than full-training in some RSSC tasks. It is worth noticing that possible factors, including dataset, initialization, optimization and domain gap are taken into consideration, and more enlightening conclusions are given through this evaluation.