U-Net Ensemble for Semantic and Height Estimation Using Coarse-Map Initialization

Single View Geocentric Pose in the Wild

“…(2) To improve prediction of taller building heights, we propose a novel strategy for fast augmentations to synthetically increase the heights of objects by inverting geocentric pose vector fields. (3) We outperform state of the art for height prediction [38,29,27,19,43] and geocentric pose [6] and demonstrate accurate predictions even for orthorectified images that violate our affine assumptions. (4) We present the first demonstration of supervising this task without lidar, using only geometry derived from images that can be produced anywhere on Earth.…”

“…Other features such as trees and buildings have known distributions of physically plausible heights. Kunwar [19] and Zheng et al [43] leveraged semantic cues as priors for height prediction to win the 2019 Data Fusion Contest (DFC19) singleview semantic 3D challenge track [20]. Srivistava et al [38] proposed to learn semantics and height jointly with a multi-task deep network.…”

“…For fair compari- son with [27,29,38,6], we retrained and tested our models and [6] using the custom DFC19 train and test sets used in [27]. For comparison with [19,43,6], we tested against the DFC19 test set used for [20]. Results are reported in Tab.…”

Disentangled Latent Transformer for Interpretable Monocular Height Estimation

“…Besides, Conditional generative adversarial network (cGAN) [15] was proposed to frame height estimation as an image translation task. Kunwar et al [16] exploited semantic labels as priors to enhance the performance of height estimation on the large-scale Urban Semantic 3D (US3D) dataset [17]. Xiong et al [18] designed and constructed a large-scale benchmark dataset for cross-dataset transfer learning on the height estimation task, which includes a large-scale synthetic dataset and several real-world datasets.…”

“…Objects of different semantic types usually have distinct height attributes. Thus, the geometric information in height maps should have high correlation with the semantic information [16]. Inspired by this, we choose to examine the learned interior activations to find human-understandable representations.…”