Object Pose Estimation · 2025 · ICCV · CCF A

MixRI: Mixing Features of Reference Images for Novel Object Pose Estimation

The method directly fuses features and occlusion states for the same 3D point across reference views, then recovers pose using RANSAC-PnP. It aims to reduce reference counts, memory, and inference costs for BOP novel-object estimation.

Paper Project

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Unseen objects

Input modalities

RGBrendered reference images

Requirements

CAD modelcamera intrinsicsrendered reference image bankobject crop/detectionRANSAC-PnP

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

Authors: Xinhang Liu, Jiawei Shi, Zheng Dang, Yuchao Dai
Affiliations: Northwestern Polytechnical University / Shaanxi Key Laboratory of Information Acquisition and Processing / EPFL CVLab
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 · Unseen objects
Datasets: BOP core datasets GSO-Datasets
Evaluation metrics: BOP AR VSD MSSD MSPD
Method tags: Novel-object Pose Estimation Template Matching & Render Retrieval 2D–3D Correspondences Keypoints + PnP Absolute Pose Estimation
Requirements: CAD modelcamera intrinsicsrendered reference image bankobject crop/detectionRANSAC-PnP

Reading Notes

Overview / Deep Dive

Available src/paper-notes/mixri-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.

Paper Reading Notes (Coarse): MixRI: Mixing Features of Reference Images for Novel Object Pose Estimation

1. Basic information

Year / venue: 2025 / ICCV; reading mode: coarse.
Task: CAD-based novel-object 6D pose from RGB and multiple rendered references.
Paper: ICCV Open Access; project: MixRI.

2. Positioning in one sentence

Instead of retrieving one nearest template, MixRI aggregates each 3D point’s evidence across reference views and directly localizes it in the query. RANSAC-PnP then solves novel-object pose with fewer references and lower caching costs.

2.1 Guide to the paper’s figures and tables

Figure / table	Focus
Figure 2	Top: retrieval-then-pose baseline; bottom: all-reference point matching.

MixRI Figure 2: comparison of reference-image pipeline — Original Figure 2: MixRI versus retrieval-based pipelines. Source: ICCV Open Access. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

3. Method overview

Sample known 3D points across reference views, aggregate per-point multi-view features, and explicitly predict query occlusion.
View-Aggregated Point Matching predicts query 2D points; 2D–3D correspondences and RANSAC-PnP recover pose.
Train on synthetic GSO; new objects need rendered CAD references, without object-specific fine-tuning.

4. Experimental overview

Report AR, network size, reference count, and speed over seven BOP core datasets.
Author claim: 12 or 24 references retain competitive AR against large-template methods while reducing memory and inference time.
My analysis: explicitly optimizing reference count suits edge budgets, though CAD, detection, and rendering costs remain.

5. Strengths, Risks, and Ideas

Strengths: avoids single-template retrieval failures, retains interpretable PnP, and uses multi-view evidence.
Risks: CAD, intrinsics, and crops remain required; occlusion-label / simulation quality affects matching.
Idea: use a 3D point’s visibility set as the matching unit instead of retrieving a reference view first.

Available src/paper-notes/mixri-iccv2025/deep.md

In-depth paper notes

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

Paper Reading Notes (Deep): MixRI: Mixing Features of Reference Images for Novel Object Pose Estimation

1. Basic information

Authors / affiliations: Xinhang Liu et al.; Northwestern Polytechnical University, Shaanxi Key Laboratory of Information Acquisition and Processing, EPFL CVLab.
Year / venue: 2025 / ICCV; reading mode: deep.
Links: ICCV Open Access / Project.

2. Positioning in one sentence

MixRI is CAD-based generalizable pose: replace query-to-nearest-template matching with fusion of all reference projections of each 3D point, establishing 2D–3D correspondences from few templates.

2.1 Guide to the paper’s figures and tables

MixRI Figure 2: multi-reference matching pipeline — Original Figure 2: retrieval + pose above, multi-reference correspondence fusion below. Source: ICCV Open Access. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

3. Task and Fair Comparison

Training: synthetic GSO-Datasets; render references with intrinsics / poses for each CAD object.
Inference: query RGB, references, and 3D point projections; predict query 2D positions / occlusion flags, then RANSAC-PnP.
Like MegaPose / GigaPose / FoundPose, test objects have CAD. Inputs differ from CAD-free BoxDreamer or single-reference RGB-D methods.

4. Method details

Module	Mechanism	Role
Reference point sampling	Projections of one 3D point across rendered views	Multi-view appearance
View-aggregated matching	Fuse all features and visibility	Avoid dependence on one retrieved template
Query point / occlusion head	Predict 2D position and invisibility	Reject occluded / out-of-frame matches
RANSAC-PnP	Use 2D–3D correspondences	Interpretable 6D pose

Author claim: only 12 references and no offline feature cache yield competitive AR across seven BOP core datasets, jointly considering reference time, memory, parameters, and FPS.
My analysis: all-reference fusion strengthens geometric evidence, but growing reference counts, attention / fusion cost, and noisy-reference robustness merit separate ablations.

5. Experimental analysis

Table 1 compares BOP AR, parameters, reference counts, and speed; Table 2 further constrains reference counts.
The main case is fewer templates, faster and smaller at comparable accuracy, matching edge library / cache budgets rather than one benchmark record.
Reproduction should align detection timing, rendering timing, pre-extracted features, and BOP object-crop sources.

6. Strengths, Limitations, and Extensions

Strengths: explicit occlusion outputs, diagnosable PnP correspondence errors, and broad engineering metrics.
Limits: CAD, masks / detection, and intrinsics remain required. Natural rather than rendered references, reflections / transparency, and severe domain gaps may differ.
Extensions: uncertainty-guided PnP, adaptive view coverage with 12 references, and foundation features with lightweight heads.

7. Research Lessons

Aggregate visible evidence per 3D point instead of selecting one best template.
An efficiency baseline for CAD-based RGB novel-object pose; CAD-free studies need separate reconstruction-based comparisons.

8. Additional Original Results

MixRI BOP results, reference-count ablations, and visualizations — Crop of original experimental pages: seven-dataset BOP results, varying reference counts, and YCB-V qualitative examples. Source: ICCV 2025 paper PDF. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

Read results with reference budgets: MixRI retains cross-dataset AR using few references rather than hundreds of cached templates. Too little coverage misses appearances, while redundant or poor views increase fusion costs. Fix counts, rendering resolution, and caching before comparing speed / size claims.

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.