[Paper Review] End-to-end Recovery of Human Shape and Pose

Angjoo Kanazawa, Michael J. Black+2 more|arXiv (Cornell University)|Dec 18, 2017

Human Pose and Action Recognition60 citations

TL;DR

Presents Human Mesh Recovery (HMR), an end-to-end system that reconstructs a full 3D SMPL mesh from a single RGB image using reprojection loss and a factorized adversarial prior, trainable with or without paired 2D-3D data, and running in real time.

ABSTRACT

We describe Human Mesh Recovery (HMR), an end-to-end framework for reconstructing a full 3D mesh of a human body from a single RGB image. In contrast to most current methods that compute 2D or 3D joint locations, we produce a richer and more useful mesh representation that is parameterized by shape and 3D joint angles. The main objective is to minimize the reprojection loss of keypoints, which allow our model to be trained using images in-the-wild that only have ground truth 2D annotations. However, the reprojection loss alone leaves the model highly under constrained. In this work we address this problem by introducing an adversary trained to tell whether a human body parameter is real or not using a large database of 3D human meshes. We show that HMR can be trained with and without using any paired 2D-to-3D supervision. We do not rely on intermediate 2D keypoint detections and infer 3D pose and shape parameters directly from image pixels. Our model runs in real-time given a bounding box containing the person. We demonstrate our approach on various images in-the-wild and out-perform previous optimization based methods that output 3D meshes and show competitive results on tasks such as 3D joint location estimation and part segmentation.

Motivation & Objective

• Motivate why single-view 3D human mesh reconstruction is valuable beyond sparse 3D joints.
• Develop an end-to-end framework that directly regresses SMPL shape and pose from image features.
• Address lack of in-the-wild 3D ground truth via a data-driven adversarial prior.
• Demonstrate real-time performance and applicability to in-the-wild images and segmentation tasks.

Proposed method

• Use SMPL to parameterize body shape and pose as 10 shape coefficients and 3K pose rotations.
• Employ an iterative error feedback loop to regress 85D SMPL-related parameters from image features.
• Train a discriminator network for a factorized adversarial prior over shape and pose to regularize outputs to the human manifold.
• Optimize a reprojection loss that enforces 2D keypoints to align with projected 3D joints, enabling weak supervision from 2D keypoint annotations only.
• Optionally incorporate ground-truth 3D losses when available (joints and SMPL parameters) for improved accuracy.
• Operate in a weakly-supervised setting and allow training without paired 3D supervision while maintaining competitive 3D joint estimation and segmentation performance.

Experimental results

Research questions

• RQ1 Can a single RGB image be used to recover a full 3D SMPL mesh without relying on 2D keypoint detections?
• RQ2 Does a data-driven adversarial prior enable realistic 3D human meshes when 3D ground truth is unavailable?
• RQ3 How does end-to-end SMPL parameter regression compare to multi-stage 3D pose estimation methods in in-the-wild images?
• RQ4 Is the approach capable of real-time inference and useful for downstream tasks like segmentation?

Key findings

• The method yields a full 3D human mesh by regressing SMPL shape and pose directly from image features.
• An iterative 3D regression with feedback improves recovery stability and accuracy.
• A factorized adversarial prior (shape and per-joint pose discriminators) regularizes outputs to lie on the human manifold, enabling training without paired 3D data.
• The approach outperforms previous SMPL-from-2D-keypoint methods on 3D joint estimation benchmarks and is competitive with methods predicting only 3D joints.
• The model runs in real time given a person bounding box and supports auxiliary tasks such as body part segmentation.
• Experiments show strong performance even without any paired 3D supervision (HMR unpaired).