RDFC-GAN: RGB-Depth Fusion CycleGAN for Indoor Depth Completion

TL;DR

This work tackles indoor depth completion where depth sensors produce large holes due to material properties and geometry. It introduces a two-branch network (MCN and RGB-Depth Fusion CycleGAN) called RDFC-GAN, which fuses an incomplete depth map with an RGB image while leveraging Manhattan-world regularities and texture-rich CycleGAN-based synthesis. The model includes a novel W-AdaIN fusion mechanism, a confidence-based fusion head, and pseudo depth maps to better simulate indoor missing patterns during training. Extensive evaluation on NYU-Depth V2 and SUN RGB-D shows state-of-the-art performance in depth values and geometry (point clouds), with demonstrated improvements for downstream tasks like 3D object detection. Overall, the approach advances indoor depth completion by effectively combining geometry-guided and texture-guided cues, yielding robust dense depth maps across diverse indoor scenes.

Abstract

Raw depth images captured in indoor scenarios frequently exhibit extensive missing values due to the inherent limitations of the sensors and environments. For example, transparent materials frequently elude detection by depth sensors; surfaces may introduce measurement inaccuracies due to their polished textures, extended distances, and oblique incidence angles from the sensor. The presence of incomplete depth maps imposes significant challenges for subsequent vision applications, prompting the development of numerous depth completion techniques to mitigate this problem. Numerous methods excel at reconstructing dense depth maps from sparse samples, but they often falter when faced with extensive contiguous regions of missing depth values, a prevalent and critical challenge in indoor environments. To overcome these challenges, we design a novel two-branch end-to-end fusion network named RDFC-GAN, which takes a pair of RGB and incomplete depth images as input to predict a dense and completed depth map. The first branch employs an encoder-decoder structure, by adhering to the Manhattan world assumption and utilizing normal maps from RGB-D information as guidance, to regress the local dense depth values from the raw depth map. The other branch applies an RGB-depth fusion CycleGAN, adept at translating RGB imagery into detailed, textured depth maps while ensuring high fidelity through cycle consistency. We fuse the two branches via adaptive fusion modules named W-AdaIN and train the model with the help of pseudo depth maps. Comprehensive evaluations on NYU-Depth V2 and SUN RGB-D datasets show that our method significantly enhances depth completion performance particularly in realistic indoor settings.

Figures (9)

• Figure 1: Showcases of the raw depth maps (top) in indoor scenarios collected by different sensors from the SUN RGB-D dataset song2015sun and the corresponding depth completion results (bottom) of our method.
• Figure 2: Depth data visualizations of indoor RGB-Depth sensor data (top, NYU-Depth V2) and outdoor Lidar scan data (bottom, KITTI). The downsampled data ($\mathcal{T^*}$) is 500 pixels randomly and uniformly sampled from the ground-truth (GT) depth data ($\mathcal{T}$), which contains ground truth depth values (e.g., in the red box) that do not exist in the raw depth data ($\mathcal{R}$).
• Figure 3: The overview of the proposed end-to-end depth completion method (RDFC-GAN). Compared to the preliminary model RDF-GAN wang2022rgb, the Manhattan normal module and the CycleGAN are the main structural improvements in RDFC-GAN.
• Figure 4: An Illustration of the Manhattan normal module in the Manhattan-Constraint network (MCN).
• Figure 5: An Illustration of the encoder-decoder structure in the two branches.