BGDNet: Background-guided Indoor Panorama Depth Estimation

Jiajing Chen, Zhiqiang Wan, Manjunath Narayana, Yuguang Li, Will Hutchcroft, Senem Velipasalar, Sing Bing Kang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 1272-1281

Depth estimation from single perspective image has received significant attention in the past decade whereas the same task applied to single panoramic image remains comparatively under-explored. Most existing depth estimation models for panoramic images imitate models proposed for perspective images which take RGB images as input and output depth directly. However as demonstrated by our experiments model performance drops significantly when the training and testing datasets greatly differ since they overfit the training data. To address this issue we propose a novel method referred to as the Background-guided Network (BGDNet) for more robust and accurate depth estimation from indoor panoramic images. Different from existing models our proposed BGDNet first infers the background depth namely from walls floor and ceiling via background masks room layout and camera model. The background depth is then used to guide and improve the output foreground depth. We perform within dataset as well as cross-domain experiments on two benchmark datasets. The results show that BGDNet outperforms the state-of-the-art baselines and is more robust to overfitting issues with superior generalization across datasets