We present a novel 3-D instance segmentation framework for multiview stereo (MVS) buildings in urban scenes. Unlike existing works focusing on semantic segmentation of urban scenes, the emphasis of this work lies in detecting and segmenting 3-D building instances even if they are attached and embedded in a large and imprecise 3-D surface model. Multiview red green blue (RGB) images are first enhanced to RGB height (RGBH) images by adding a heightmap and are segmented to obtain all roof instances using a fine-tuned 2-D instance segmentation neural network. Instance masks from different multiview images are then clustered into global masks. Our mask clustering accounts for spatial occlusion and overlapping, which can eliminate segmentation ambiguities among multiview images. Based on these global masks, 3-D roof instances are segmented out by mask back-projections and extended to the entire building instances through a Markov random field optimization. A new dataset that contains instance-level annotation for both 3-D urban scenes (roofs and buildings) and drone images (roofs) is provided. To the best of our knowledge, it is the first outdoor dataset dedicated to 3-D instance segmentation with much more annotations of attached 3-D buildings than existing datasets. 1 Quantitative evaluations and ablation studies have shown the effectiveness of all major steps and the advantages of our multiview framework over the orthophoto-based method.
Depth maps are acquirable and irreplaceable geometric information that significantly enhances traditional color images. RGB and Depth (RGBD) images have been widely used in various image analysis applications, but they are still very limited due to challenges from different modalities and misalignment between color and depth. In this paper, a Fully Aligned Fusion Network (FAFNet) for RGBD semantic segmentation is presented. To improve cross‐modality fusion, a new RGBD fusion block is proposed, features from color images and depth maps are first fused by an attention cross fusion module and then aligned by a semantic flow. A multi‐layer structure is also designed to hierarchically utilize the RGBD fusion block, which not only eases issues of low resolution and noises for depth maps but also reduces the loss of semantic features in the upsampling process. Quantitative and qualitative evaluations on both the NYU‐Depth V2 and the SUN RGB‐D dataset demonstrate that the FAFNet model outperforms state‐of‐the‐art RGBD semantic segmentation methods.