[Paper Review] Pose from Shape: Deep Pose Estimation for Arbitrary 3D Objects

Yang Xiao, Xuchong Qiu+3 more|arXiv (Cornell University)|Jun 12, 2019

Robot Manipulation and Learning59 references36 citations

TL;DR

A generic, category-agnostic pose estimation method that conditions 3D object pose on a given 3D model, enabling pose estimation for unseen object categories without additional training. The approach improves performance on standard benchmarks and demonstrates strong generalization to novel objects and datasets.

ABSTRACT

Most deep pose estimation methods need to be trained for specific object instances or categories. In this work we propose a completely generic deep pose estimation approach, which does not require the network to have been trained on relevant categories, nor objects in a category to have a canonical pose. We believe this is a crucial step to design robotic systems that can interact with new objects in the wild not belonging to a predefined category. Our main insight is to dynamically condition pose estimation with a representation of the 3D shape of the target object. More precisely, we train a Convolutional Neural Network that takes as input both a test image and a 3D model, and outputs the relative 3D pose of the object in the input image with respect to the 3D model. We demonstrate that our method boosts performances for supervised category pose estimation on standard benchmarks, namely Pascal3D+, ObjectNet3D and Pix3D, on which we provide results superior to the state of the art. More importantly, we show that our network trained on everyday man-made objects from ShapeNet generalizes without any additional training to completely new types of 3D objects by providing results on the LINEMOD dataset as well as on natural entities such as animals from ImageNet.

Motivation & Objective

• Motivate robust pose estimation for objects outside predefined categories or instances (in the wild).
• Propose a deep network that conditions pose estimation on a 3D model of the target object.
• Demonstrate that shape-conditioned pose estimation improves accuracy on known categories and generalizes to novel objects.
• Show that both point-cloud and multi-view renderings of the 3D shape can be used to encode shape information for pose prediction.

Proposed method

• Two-branch network processes: (1) RGB image through a CNN (ResNet-18) and (2) 3D shape through either PointNet or multi-view rendered images.
• A mixed classification-and-regression loss predicts Euler-angle bins and intra-bin offsets for azimuth, elevation, and in-plane rotation.
• Angles are discretized into L_theta bins with corresponding classification scores and regression offsets (Huber loss).
• Data augmentation includes shape-rotation perturbations to reduce overfitting to canonical orientations.
• Training uses Adam with staged learning rates; synthetic ShapeNet data used for training with SUN397 backgrounds; testing on Pascal3D+, ObjectNet3D, Pix3D, and LINEMOD.
• Shape encoders: (a) PointNet for point clouds; (b) multi-view CNN using rendered views around the object; weights shared across viewpoints.

Experimental results

Research questions

• RQ1 Can a deep pose estimator learn category-free viewpoint estimation conditioned on a 3D object model?
• RQ2 Does incorporating exact or approximate 3D shape information boost pose estimation performance for known categories?
• RQ3 How well does the method generalize to novel categories and completely unseen object types?
• RQ4 What is the impact of using multi-view shape representations versus single-view or point-cloud encodings?

Key findings

• Using 3D shape information (point cloud or multi-view renderings) significantly improves pose estimation over a shape-less baseline across datasets.
• Multi-view representations generally outperform point-cloud encodings for the shape input.
• The method achieves competitive or superior results on Pascal3D+, ObjectNet3D, and Pix3D, even when trained only on synthetic data.
• The approach provides meaningful coarse pose estimates on LINEMOD without object-specific training, enabling effective downstream refinement (e.g., DeepIM).
• Randomizing object shape orientation during training reduces overfitting to a canonical pose and improves robustness to unseen shapes.