Object Pose Estimation · 2025 · ICCV · CCF A

Ultra-Precision 6DoF Pose Estimation Using 2-D Interpolated Discrete Fourier Transform

A 2D interpolated DFT estimates 6-DoF pose from the phase and frequency of periodic patterns, with a mathematical mapping from image parameters to pose. It targets controlled industrial measurements at nanometer translation and microradian rotation scales.

Paper

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Seen objects

Input modalities

industrial RGB cameraperiodic pattern image

Requirements

calibrated industrial cameraperiodic photometric patternorthographic or quasi-orthographic imaging

Full paper metadata Authors, affiliations, research area, datasets, metrics, method tags, and reading status Expand

Authors: Guowei Shi, Zian Mao, Peisen Huang
Affiliations: UM-SJTU Joint Institute, Shanghai Jiao Tong University
Conference / journal: Conference · ICCV · CCF A
Research area: Object Pose Estimation
Code availability: Unavailable / not yet recorded
Reading status: Overview:Available Deep Dive:Available
Task setting: Instance-level · Seen objects
Datasets: periodic-pattern simulation industrial camera measurement setup
Evaluation metrics: translation RMSE rotation RMSE nanometer accuracy microradian accuracy
Method tags: Instance-level Pose Estimation Synthetic Data & Foundation Models Absolute Pose Estimation
Requirements: calibrated industrial cameraperiodic photometric patternorthographic or quasi-orthographic imaging

Reading Notes

Overview / Deep Dive

Available src/paper-notes/ultra-precision-6dof-iccv2025/coarse.md

Paper overview notes

Quickly assess the paper's scope, problem setting, core method, experimental findings, and whether to read it in depth.

Overview Notes: Ultra-Precision 6DoF Pose Estimation Using 2-D Interpolated Discrete Fourier Transform

1. Basic information

Year / venue: 2025 / ICCV; reading mode: coarse.
Task: ultra-precise 6DoF measurement using periodic patterns and industrial cameras.
Original paper: ICCV Open Access.

2. Positioning in one sentence

Rather than learning general object recognition, this paper uses 2D-IpDFT to precisely estimate periodic-pattern frequencies and phases, then recovers 6DoF pose through an imaging / pattern model for controlled metrology applications such as lithography and nanomanipulation.

2.1 Guide to the paper’s figures and tables

Figure / table	What to examine
Figure 1	A measurement loop with a fixed industrial camera, an optical periodic pattern, and a six-axis stage; this explains why direct BOP comparisons are inappropriate.

Ultra-Precision 6DoF Figure 1: periodic pattern measurement setup — Original Figure 1: periodic-pattern 6DoF measurement setup. Source: ICCV Open Access. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Method overview

A 2D interpolated DFT accounts for spectral leakage from positive / negative frequencies when estimating frequency and phase.
Analytical relationships connect frequencies and phases to translation / rotation under orthographic and quasi-orthographic camera models.
Experiments use an industrial camera and photoetched checkerboard / grid patterns; the object prior is therefore a high-quality controlled pattern, rather than natural-object shape.

4. Experimental overview

Simulations analyze SNR, visible pattern periods, and defocus, followed by validation on low-cost hardware.
Authors’ claim: nanometer-level translation and microradian-level rotation accuracy, with greater robustness to defocus / spectral leakage.
My assessment: valuable for precision engineering, but applicability depends on optics, calibration, and pattern manufacturing; it does not establish performance on occluded, cluttered, or textureless natural objects.

5. Strengths, risks, and insights

Strength: the analytical model corresponds directly to experimental error sources, with metrics closer to the application than generic vision benchmarks.
Risk: controlled patterns, near-orthographic imaging, and stable illumination are strong prerequisites.
Insight: evaluate 6D pose at application-specific precision scales; micro / nanoscale metrology and robotic grasping should not rely only on the same benchmark terminology.

Available src/paper-notes/ultra-precision-6dof-iccv2025/deep.md

In-depth paper notes

Support reproduction, comparison, related-work writing, and identifying opportunities for further research.

Deep Reading Notes: Ultra-Precision 6DoF Pose Estimation Using 2-D Interpolated Discrete Fourier Transform

1. Basic information

Authors / affiliation: Guowei Shi, Zian Mao, Peisen Huang; University of Michigan–Shanghai Jiao Tong University Joint Institute.
Year / venue: 2025 / ICCV; reading mode: deep.
Original paper: ICCV Open Access.

2. Positioning in one sentence

To address small translation / rotation biases caused by periodic-pattern spectral leakage under near-orthographic industrial imaging, 2D-IpDFT estimates frequency and phase more precisely, then uses analytical geometry to recover 6DoF at nanometer / microradian scales.

2.1 Guide to the paper’s figures and tables

Ultra-Precision 6DoF Figure 1: industrial periodic-pattern setup — Original Figure 1: a measurement system comprising an industrial camera, periodic pattern, and six-axis nanopositioning stage. Source: ICCV Open Access. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Problem setting and boundaries

Input: optical periodic patterns such as checkerboards / grids, imaged by a fixed, low-distortion industrial camera.
Output: the pattern’s $t_x,t_y,t_z,\alpha,\beta,\gamma$ relative to the camera.
Camera model: orthographic or quasi-orthographic; the paper studies frequency / phase estimation bias with few visible periods, defocus, and varying SNR.
It shares the $SE(3)$ output with natural-object 6D pose, while addressing different challenges from recognition, occlusion, and category generalization.

4. Method details

Module	Input / output	Role
2D DFT	Pattern image → spectral peaks	Convert spatial periodicity into frequency / phase observables
2D-IpDFT	Peak neighborhood → corrected frequency / phase	Account for positive and negative frequencies to reduce spectral leakage
Camera-pattern model	Frequency / phase and camera parameters	Analytically map image parameters to 6DoF
Feedback control	Continuously estimated poses	Demonstrate stage tracking driven by the measurements

Authors’ claim: greater robustness to SNR, pattern period, and defocus than frequency-domain baselines such as QSE, linear regression, and zero-padding.
The key is whether the estimator removes systematic bias from nearby spectral lines in a finite observation window, rather than network capacity.

5. Experimental analysis

Simulations vary SNR, pattern periods, and Gaussian defocus, measuring translation / rotation RMSE along each axis.
The physical system uses an industrial camera and photoetched checkerboard, reporting nanometer-level translation and microradian-level rotation; later figures also demonstrate vision-based feedback control.
My assessment: these experiments adequately support the target application, but their accuracy should not be compared directly with RGB object-pose benchmarks lacking the same hardware / texture controls.

6. Critical analysis and reproduction

Strengths: the system model, error sources, simulation variables, and hardware validation align well, making the approach highly interpretable for precision metrology.
Limitations: sensitivity to pattern-manufacturing accuracy, lens distortion, camera-noise models, and external vibration; objects without the patterns are outside its scope.
Reproduction priorities: camera calibration, pattern pitch, DFT window / interpolation parameters, stage ground truth, and thermal-drift control.
Extensions: fusion with depth / phase cameras; estimator covariance in the control loop; larger working distances and nonideal illumination.

7. Research insights

6D pose is not always a recognition problem: when objects / patterns can be designed, imaging models and signal estimation may enable substantially higher precision than general learning methods.

8. Additional original experimental figures and tables

2D-IpDFT trajectory RMSE, exposure-time, pattern-period, and defocus experiments — Original Tables 4–7 and Figures 9–10: helical-trajectory translation RMSE; circular-trajectory RMSE under different exposures / pattern-period counts / defocus levels; and a bar plot of in-plane trajectory RMSE versus defocus distance with sample pattern images. This is neither an outlier experiment nor an error-distribution histogram. Source: ICCV 2025 paper PDF. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

These tables and bars examine a metrology error budget, comparing trajectory-measurement algorithms and isolating exposure time, visible periods, and defocus. The bar plot’s horizontal axis is defocus distance, so it cannot establish whether the error distribution has a long tail. Accuracy varies by degree of freedom and condition; in-plane precision cannot be generalized to every translation axis. Nanometer / microradian conclusions apply only with controlled patterns, near-orthographic imaging, and a high-precision stage, not ordinary RGB views of textureless natural objects.

Related papers

2026 · CVPR

3PT

3PT unifies detection, segmentation, and 6-DoF pose with two Transformers for 2D / 3D object-centric understanding. Continuous refinement of object representations and rendered-feature alignment support zero-shot cross-object perception on BOP-H3, BOP-Industrial, and BOP-Classic.

2026 · AAAI

6DAttack

Studies backdoor attacks on PnP-based and end-to-end 6-DoF pose pipelines. Structurally 3D triggers and modified pose labels are injected into training data so that models behave normally on clean inputs but produce controlled incorrect poses when triggered.

2026 · ICRA

Active Tactile Exploration

A violation-implicit loss incorporates physical-constraint violations into joint estimation of freely moving unknown convex rigid-body trajectories and cuboid / 20-vertex polytope shapes. Observed / Fisher information at a local MLE estimates EIG to select the next touch. In simulation, the EIG strategy yields stable errors below 4 mm after 5 cuboid actions; real Trifinger results at step 5 substantially outperform random selection. Convex geometry, local optima, multiple initializations, and tactile / proprioceptive errors limit the method.