Monocular Human-Object Reconstruction in the Wild

TL;DR

This work tackles monocular human-object reconstruction by learning a 3D spatial prior directly from in-the-wild 2D images. It introduces a normalizing-flow-based prior over 2D human-object keypoint layouts and viewports, conditioned on image features, and integrates it into a two-stage post-optimization to refine 3D pose without 3D supervision. A new WildHOI dataset is collected and used alongside BEHAVE to demonstrate that 2D-derived priors can achieve near 3D-supervised performance indoors and superior generalization outdoors, improving cross-view coherence and handling non-contact interactions. The approach offers a scalable, data-driven pathway to robust HOI reconstruction in diverse real-world settings, with potential impact on AR, robotics, and human activity understanding.

Abstract

Learning the prior knowledge of the 3D human-object spatial relation is crucial for reconstructing human-object interaction from images and understanding how humans interact with objects in 3D space. Previous works learn this prior from datasets collected in controlled environments, but due to the diversity of domains, they struggle to generalize to real-world scenarios. To overcome this limitation, we present a 2D-supervised method that learns the 3D human-object spatial relation prior purely from 2D images in the wild. Our method utilizes a flow-based neural network to learn the prior distribution of the 2D human-object keypoint layout and viewports for each image in the dataset. The effectiveness of the prior learned from 2D images is demonstrated on the human-object reconstruction task by applying the prior to tune the relative pose between the human and the object during the post-optimization stage. To validate and benchmark our method on in-the-wild images, we collect the WildHOI dataset from the YouTube website, which consists of various interactions with 8 objects in real-world scenarios. We conduct the experiments on the indoor BEHAVE dataset and the outdoor WildHOI dataset. The results show that our method achieves almost comparable performance with fully 3D supervised methods on the BEHAVE dataset, even if we have only utilized the 2D layout information, and outperforms previous methods in terms of generality and interaction diversity on in-the-wild images.

Figures (3)

• Figure 1: In this work, we aim at learning instance-level human-object spatial relation prior from unlimited images in the wild. To accomplish this, we utilize the normalizing flow to learn the view distribution and the 2D human-object keypoints layout on each image plane. The spatial relation prior is then applied to real-world images under the monocular human-object reconstruction setting.
• Figure 2: The main pipeline of our method. We utilize the normalizing flow to learn the distribution of the 2D human-object keypoints in each image plane from vast images in the wild. The normalizing flow takes the input image $\mathbf{I}$ as the condition to transform the noize $\mathbf{z}$ from Gaussian distribution to the 2.5D keypoints $\mathbf{X}_{\text{2.5D}}$ which is intermediate representation combining the view pose $\rho$ and the 2D human-object keypoint layout $\Pi_\rho(\mathbf{X}_{\text{3D}})$. To train this conditioned normalizing flow, we collect a bunch of images from the Internet and group these images together based on the geometry consistency of the 2D human-object keypoints in each view. Then we incorporate the prior learned from 2D images into the post-optimization process. In the post-optimization stage, we project the 3D human-object keypoints onto different image planes of the virtual cameras to ensure the reconstructed results seem coherently observed from other views. Besides, we use the mean occlusion maps that are obtained by averaging the occlusion maps in the images to compute the contact loss. Our method is supervised without using any 3D annotations or commonsense knowledge of the spatial relation between the human and the object.
• Figure 3: The qualitative results on WildHOI dataset.