Prior-free relative 6D pose estimation
of multiple object instances

Behdad Khodabandehloo1,2, Andrea Caraffa1, Davide Boscaini1, Fabio Poiesi1
1Fondazione Bruno Kessler, Trento, Italy
2University of Trento, Trento, Italy
{bkhodabandehloo, acaraffa, dboscaini, poiesi}@fbk.eu
Paper Code (Coming soon)
Single-reference setting Prior-free setting (ours)
Comparison between single-reference and prior-free relative 6D pose estimation
PROSE estimates how one instance of an object is positioned and oriented relative to another. Existing methods compare the target object (query, red) in the input image I against an anchor (green) supplied in a separate reference image J (left). PROSE selects both the query and anchor directly from the same RGB-D image I (right), so it requires no CAD model, template, or external reference image.

Abstract

Object 6D pose estimation formulations have progressively reduced reliance on object-specific priors, evolving from explicit 3D models to multi-view object captures to single reference images. We take this progression to its extreme by introducing prior-free relative 6D pose estimation, which lifts the assumption of knowing which object is to be posed within the scene. This novel setting aims to estimate the relative poses of multiple instances of an unknown object within the same image, without requiring CAD models, templates, or reference images. We solve this by formulating a novel method (PROSE) that finds coarse correspondences between object instances using multimodal foundation features, thus requiring no training. We refine these correspondences by imposing cycle consistency across tuples of instances, and leverage the resulting globally consistent correspondences to estimate the relative 6D pose between any pair of instances. To enable systematic evaluation, we design a novel benchmark (PRENCH) built from three multi-instance BOP datasets and enriched with task-specific metadata. PROSE consistently outperforms baselines obtained by adapting state-of-the-art single-image methods to the proposed setting, while requiring neither task-specific supervision nor additional learned components.

Cycle-consistent Correspondence Refinement

Before-and-after comparison showing conflicting object matches before refinement and consistent matches afterward
This figure shows how PROSE makes matches between multiple instances of the same object more reliable. On the left, each pair of objects is matched separately, which can produce conflicting results. On the right, PROSE considers all the objects together and adjusts the matches until they agree.

Quantitative Results

Quantitative comparison on PRENCH. Higher average recall (AR) and ADD(S) scores indicate better performance, while lower rotation error (RE, measured in degrees) and translation error (TE, measured in millimeters) are better. PROSE is highlighted, and Improvement reports its margin over the strongest baseline.
Method IC-BIN IC-MI XYZ-IBD
AR ↑ ADD(S) ↑ RE (°) ↓ TE (mm) ↓ AR ↑ ADD(S) ↑ RE (°) ↓ TE (mm) ↓ AR ↑ ADD(S) ↑ RE (°) ↓ TE (mm) ↓
One2Any 17.7 61.0 70.5 26.9 47.7 98.0 47.2 12.3 9.2 49.2 85.0 33.2
ConceptPose 18.8 58.8 76.2 36.6 50.8 80.2 51.1 29.1 36.1 70.1 64.6 24.4
PROSE (ours) 38.0 83.3 61.3 16.8 70.3 100 30.4 4.6 55.8 92.1 28.6 9.1
Improvement +19.2 +22.3 −9.2 −10.1 +19.5 +2.0 −16.8 −7.7 +19.7 +22.0 −36.0 −15.3

Qualitative Results

Input image
One2Any
ConceptPose
PROSE (ours)
Qualitative comparison of input images, One2Any, ConceptPose, and PROSE on five PRENCH scenes
Qualitative results on PRENCH. Green borders denote the ground-truth 6D poses, while red borders show each method's predictions. Accurate poses show good alignment between the two colored borders. The rows show IC-BIN coffee cups and juice boxes, XYZ-IBD metallic gears and brackets, and IC-MI milk bottles.

Discussion

Across all three PRENCH datasets, PROSE achieves the strongest performance on every reported metric, improving average recall by about 19 points over the strongest competing method while also reducing rotation and translation errors. This consistent improvement is important because the datasets cover different challenges, from densely packed everyday objects affected by self-occlusion to industrial components with limited visual texture. The qualitative results support the numerical findings: PROSE's predicted poses align more closely with the ground truth, while One2Any and ConceptPose often struggle when objects overlap or appearance provides too little information. These gains come from combining complementary semantic and geometric features and then refining the initial matches jointly across all visible object instances. By enforcing cycle consistency, PROSE can correct conflicting pairwise matches and use the instances in the scene as mutual references. The method is training-free and requires no CAD model, template, external reference image, or additional task-specific learned component, reducing the setup needed for new objects. Its scope is relative pose estimation from RGB-D observations containing multiple visible instances, so it does not replace absolute-pose approaches when only one instance or no depth information is available.

Key Takeaways

  • PROSE consistently achieves higher recall and lower pose errors across all three datasets.
  • Its largest pose-error reductions occur on challenging industrial objects with limited visual texture.
  • It removes the need for object-specific references and training, provided that multiple instances and depth information are available.

Citation

If you find this work useful, please cite:

@misc{khodabandehloo2026prior,
  title         = {Prior-free Relative 6D Pose Estimation of Multiple Object Instances},
  author        = {Khodabandehloo, Behdad and Caraffa, Andrea and
                   Boscaini, Davide and Poiesi, Fabio},
  year          = {2026},
  eprint        = {2609.08949},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.08949}
}