Scene and Human in One World: Reconstruction in a Feedforward Pass

A new computer-vision framework called SHOW can reconstruct both humans and their surrounding scenes in a unified metric space from a single moving camera, according to a preprint posted to arXiv on June 26, 2026 [1]. The method addresses long-standing challenges of scale ambiguity and human-scene misalignment that have hindered monocular dynamic-scene reconstruction [1]. The framework, whose name stands for Scene and Human in One World, performs the reconstruction in a single feedforward pass rather than treating human mesh recovery and scene reconstruction as separate pipelines [1]. The authors argue that parametric human models supply semantic structure and metric-scale priors, while scene geometry provides spatial context for localizing and aligning the human figure [1]. SHOW injects those human-derived priors into normalized point-map prediction, enabling metric-scale scene reconstruction from inherently scale-ambiguous monocular input [1]. In turn, the recovered scene geometry constrains human mesh estimation, encouraging spatially consistent human placement [1]. The model learns both human-aware geometric features and geometry-constrained human features through joint training [1]. Monocular reconstruction of dynamic scenes has historically been difficult because a single camera cannot natively resolve absolute scale, and occlusions frequently break the correspondence between the person and the background [1]. SHOW incorporates a promptable masking mechanism that lets the system flexibly select a target human while suppressing background distractions and occlusion interference [1]. The preprint reports that extensive experiments show improvements in metric-scale consistency, human-scene alignment, and reconstruction accuracy under challenging camera motion, occlusion, and cluttered backgrounds [1]. The work arrives during a period of rapid architectural change in computer vision. Convolutional neural networks, which learn features through filter optimization, were long the de-facto standard for image and video tasks [2]. More recently, transformer architectures—first described in 2017—have become dominant for many sequence-processing problems because of their superior handling of long-range dependencies [3][4]. The SHOW authors do not specify in the abstract which backbone architecture they employ, but the emphasis on a feedforward pass and joint feature learning is consistent with contemporary hybrid designs that blend convolutional and attention-based components [1][2][3]. The paper appeared on arXiv, the open-access repository that has hosted more than two million e-prints since its launch in 1991 and now receives roughly 24,000 submissions per month [8]. The abstract page includes links to community-built tools under the arXivLabs framework, a program launched in 2020 that lets third-party developers offer bibliographic explorers, code finders, and recommender systems directly on article pages [5][6][7].