Despite significant progress in RGB-D based road segmentation in recent years, the latest methods cannot achieve both state-of-the-art accuracy and real time due to the high-performance reliance on heavy structures. We argue that this reliance is due to unsuitable multimodal fusion. To be specific, RGB and depth data in road scenes are each sensitive to different regions, but current RGB-D based road segmentation methods generally combine features within sensitive regions which preserves false road representation from one of the data. Based on such findings, we design an Evidence-based Road Segmentation Method (Evi-RoadSeg), which incorporates prior knowledge of the modal-specific characteristics. Firstly, we abandon the cross-modal fusion operation commonly used in existing multimodal based methods. Instead, we collect the road evidence from RGB and depth inputs separately via two low-latency subnetworks, and fuse the road representation of the two subnetworks by taking both modalities’ evidence as a measure of confidence. Secondly, we propose an RGB-D data augmentation scheme tailored to road scenes to enhance the unique properties of RGB and depth data. It facilitates learning by adding more sensitive regions to the samples. Finally, the proposed method is evaluated on the widely used KITTI-road, ORFD, and R2D datasets. Our method achieves state-of-the-art accuracy at over 70 FPS, 5$\times$faster than comparable RGB-D methods. Furthermore, extensive experiments illustrate that our method can be deployed on a Jetson Nano 2GB with a speed of 8$+$FPS. The code will be released in https://github.com/xuefeng-cvr/Evi-RoadSeg.

The high performance of RGB-D based road segmentation methods contrasts with their rare application in commercial autonomous driving, which is owing to two reasons: 1) the prior methods cannot achieve high inference speed and high accuracy in both ways; 2) the different properties of RGB and depth data are not well-exploited, limiting the reliability of predicted road. In this paper, based on the evidence theory, an uncertainty-aware symmetric network (USNet) is proposed to achieve a trade-off between speed and accuracy by fully fusing RGB and depth data. Firstly, cross-modal feature fusion operations, which are indispensable in the prior RGB-D based methods, are abandoned. We instead separately adopt two light-weight subnetworks to learn road representations from RGB and depth inputs. The light-weight structure guarantees the real-time inference of our method. Moreover, a multi-scale evidence collection (MEC) module is designed to collect evidence in multiple scales for each modality, which provides sufficient evidence for pixel class determination. Finally, in uncertainty-aware fusion (UAF) module, the uncertainty of each modality is perceived to guide the fusion of the two sub-networks. Experimental results demonstrate that our method achieves a state-of-the-art accuracy with real-time inference speed of$43+$FPS. The source code is available at https://github.com/morancyc/USNet.

The performance of monocular depth estimation generally depends on the amount of parameters and computational cost. It leads to a large accuracy contrast between light-weight networks and heavy-weight networks, which limits their application in the real world. In this paper, we model the majority of accuracy contrast between them as the difference of depth distribution, which we call 'Distribution drift'. To this end, a distribution alignment network (DANet) is proposed. We firstly design a pyramid scene transformer (PST) module to capture inter-region interaction in multiple scales. By perceiving the difference of depth features between every two regions, DANet tends to predict a reasonable scene structure, which fits the shape of distribution to ground truth. Then, we propose a local-global optimization (LGO) scheme to realize the supervision of global range of scene depth. Thanks to the alignment of depth distribution shape and scene depth range, DANet sharply alleviates the distribution drift, and achieves a comparable performance with prior heavy-weight methods, but uses only 1% floating-point operations per second (FLOPs) of them. The experiments on two datasets, namely the widely used NYUDv2 dataset and the more challenging iBims-1 dataset, demonstrate the effectiveness of our method. The source code is available at https://github.com/YiLiM1/DANet.