Pose Estimation for Textureless Objects via Illumination Invariant Tree

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing pose estimation systems struggle to accurately determine the location and positioning of textureless objects, which lack surface features necessary for estimation, hindering robotic devices' ability to grasp and manipulate them without mishandling or damaging.

Innovation Solution

A system that aligns a 3D model of the textureless object to a single image by densely sampling the space of poses, using a multi-level illumination invariant tree structure to leverage dynamic programming for simultaneous localization of patches at various levels, enabling precise alignment and invariance to illumination.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If histograms of gradients are used to provide invariance to low-level features, then robustness to illumination changes is improved, but measurement precision of pose estimation deteriorates

Engineering Contradiction:
Improverobustness to illumination changesVSAvoidpose estimation precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent segments the pose estimation process into two distinct stages: coarse alignment using illumination-invariant features (histograms of gradients) to achieve robustness, and fine alignment using raw image features to achieve precision. This segmentation allows each stage to optimize for its specific goal without compromising the other.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary coarse alignment using illumination-invariant histograms of gradients before conducting fine alignment. This preliminary action establishes a reasonable initial pose estimate that is robust to illumination changes, which then serves as the starting point for the precision-oriented fine alignment stage.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If 3D models are aligned to images using low-level primitives, then alignment speed is improved, but manufacturing precision of pose estimation deteriorates for textureless objects

Engineering Contradiction:
Improvealignment speedVSAvoidpose estimation precision for textureless objects
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent segments the alignment process into coarse alignment (using low-level primitives for speed) and fine alignment (using detailed image features for precision). This segmentation enables the system to achieve both speed and precision, particularly for textureless objects where the fine alignment stage can recover details that low-level primitives miss.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs partial alignment using low-level primitives to achieve a reasonable initial pose estimate quickly, then performs additional fine alignment to achieve the required precision. This partial action approach allows the system to benefit from both the speed of primitive-based alignment and the precision of detailed feature matching.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If exhaustive search of pose space is performed, then measurement precision is improved, but loss of time increases

Engineering Contradiction:
Improvepose estimation precisionVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the pose search space into two stages: coarse search using illumination-invariant features to quickly eliminate incorrect poses, and fine search using detailed image features to precisely determine the final pose. This segmentation dramatically reduces the effective search space at each stage, achieving both precision and speed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary coarse pose estimation using computationally efficient illumination-invariant features before conducting the more computationally intensive fine pose estimation. This preliminary action eliminates most incorrect poses early, reducing the time required for the exhaustive fine search.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3420532B1Systems and methods for estimating pose of textureless objects
Publication Date: 2020.05.06 CARNEGIE MELLON UNIV
  • EP3420532B1 patent drawingFigure 1
  • EP3420532B1 patent drawingFigure 2A
  • EP3420532B1 patent drawingFigure 2B~2C

AI summary

Systems and methods for estimating the pose of a textureless object are disclosed. A method to estimate a pose of a textureless object includes obtaining, by a processing device, a single image of the textureless object. The pose of the textureless object can be inferred from the single image. The method further includes generating, by the processing device, a three dimensional model of the textureless object from a plurality of viewpoints and a plurality of scales obtained from image data of the textureless object, matching, by the processing device, the single image with a discretized render of the three dimensional model via a multi-level illumination invariant tree structure to obtain an alignment of the single image with the discretized render, and estimating, by the processing device, the pose of the textureless object based on the alignment.