Object Pose Estimation · 2026 · 3DV · CCF Unranked

CAOA: Completion-Assisted Object-CAD Alignment

CAOA first completes missing object geometry in scans using context-aware point completion. Symmetry encoding and a Siamese object–CAD alignment network then predict 9-DoF alignment: translation, rotation, and per-axis scale.

Paper Code

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Seen objects

Input modalities

indoor RGB-D scan / object point cloudCAD model

Requirements

instance segmentationobject point cloud with scene contextretrieved CAD model

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

Authors: Hiranya Garbha Kumar, Minhas Kamal, Balakrishnan Prabhakaran
Affiliations: University at Albany, State University of New York
Conference / journal: Conference · 3DV · CCF Unranked
Research area: Object Pose Estimation
Code availability: Code available
Reading status: Overview:Available Deep Dive:Available
Task setting: Instance-level · Seen objects
Datasets: Scan2CAD ScanNetv2 S2C-Completion SN-Indoor
Evaluation metrics: Chamfer Distance
Method tags: Instance-level Pose Estimation Template Matching & Render Retrieval 2D–3D Correspondences Pose Refinement Absolute Pose Estimation
Requirements: instance segmentationobject point cloud with scene contextretrieved CAD model

Reading Notes

Overview / Deep Dive

Available src/paper-notes/caoa-3dv2026/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 Overview: CAOA: Completion-Assisted Object-CAD Alignment

1. Basic information

Authors / affiliations: Hiranya Garbha Kumar, Minhas Kamal, Balakrishnan Prabhakaran; University at Albany.
Year / venue: 2026 / 3DV.
Paper: OpenReview; arXiv PDF.
Code: MinhasKamal/CAOA.
Reading mode: Overview.

2. Positioning in one sentence

CAOA first completes object point clouds from noisy / partial indoor scans, then passes completion and symmetry features to an object–CAD alignment network to predict translation, rotation, and per-axis scale as a 9DoF alignment.

3. Problem setting

Inputs: Segmented object point clouds from indoor RGB-D scans, nearby scene context, and a CAD model.
Output: A 9DoF pose aligning the CAD object to the real scan.
This is Scan2CAD-style semantic reconstruction rather than single-image BOP 6DoF. Scale and instance segmentation are key variables.

4. Method overview

CAPCM completes the object point cloud using the target and 100 cm of scene context.
SEM learns symmetry vectors from canonical CAD models to resolve rotation ambiguity.
OCAM encodes object / CAD point clouds with a Siamese 3D CNN and regresses $t$ , $r$ , and $s$ through separate MLP heads.

5. Experimental overview

Scan2CAD correctness requires translation ≤20 cm, rotation ≤20°, and scale error ≤20%.
Full CAOA (CAPCM+SEM) reaches class / weighted accuracy of 69.17 / 77.51 with predicted segmentation, versus 48.63 / 60.92 without completion.
Ground-truth ScanNet labels raise accuracy to 78.15 / 86.54, indicating approximately 10% performance loss from instance segmentation.
The full pipeline takes approximately 0.58 s / scene (0.56–0.59 s for 20 objects, depending on the segmentation backend).

6. Strengths, limitations, and implications

Strengths

Treats input geometry quality as a central alignment problem, supported by S2C-Completion / SN-Indoor.
Explicit symmetry and 9DoF scale handling suit real indoor scans.

Limitations / risks

Completion, SEM, and pose heads are trained separately, without fully end-to-end objectives.
Still requires reliable instance segmentation and a matching CAD model; errors in either frontend can be amplified.

7. Worth examining in depth

Point completion contributes far more than the symmetry head. A useful follow-up is letting pose uncertainty guide completion, beyond producing a visually complete point cloud.

Available src/paper-notes/caoa-3dv2026/deep.md

In-depth paper notes

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

Deep Reading Notes: CAOA: Completion-Assisted Object-CAD Alignment

1. Basic information

Authors / affiliations: Hiranya Garbha Kumar, Minhas Kamal, Balakrishnan Prabhakaran; University at Albany.
Year / venue: 2026 / 3DV.
Paper page · arXiv PDF · Code.
Reading mode: Deep Dive.

2. Positioning in one sentence

CAOA argues that incomplete scan geometry, rather than pose-regressor capacity, is often the main bottleneck in indoor object–CAD alignment. It first completes points using scene context, then injects symmetry features into 9DoF alignment.

2.1 Guide to the paper’s figures and tables

Figure / table	Reading focus
Figure 1	Dependencies between object masks, scene context, completion, symmetry, and OCAM
Figures 3–4	Raw versus completed point clouds and the Siamese alignment architecture
Tables 1–3	Full method versus no completion, no SEM, and ground-truth segmentation; examine accuracy and scene runtime together

CAOA Figure 1: completion-assisted object CAD alignment pipeline — Original paper, method-page region containing Figure 1: Instance masks and room scans lead to completion and symmetry-aware 9DoF alignment. Source: arXiv PDF. Copyright belongs to the paper's authors and publisher; reproduced here solely for paper discussion.

CAOA Table 1: Scan2CAD alignment results — Original paper, results-page region containing Table 1: Class accuracy and CAOA ablations for Scan2CAD 9DoF alignment. Source: arXiv PDF. Copyright belongs to the paper's authors and publisher; reproduced here solely for paper discussion.

3. Problem setting and boundaries

Inputs: A ScanNet-style room scan, instance segmentation mask, local object context, and a candidate CAD model.
Outputs: 9DoF alignment comprising $t\in\mathbb{R}^3$ , rotation, and per-axis scale, rather than fixed-scale BOP 6D pose.
Evaluation: Scan2CAD counts an alignment as correct only with $t\le20$ cm, $r\le20^\circ$ , and scale error $\le20\%$ .
Key assumptions: The object is already segmented and CAD retrieval / selection is correct. These prerequisites have costs within a full semantic-reconstruction system.

4. Method details

Module	Input	Output	Purpose
CAPCM	Object points + 100 cm context	Completed object point cloud	Reduce ambiguity from occlusion / scan holes using surrounding context
SEM	Canonical CAD point cloud	Symmetry vector	Explicitly inform the regressor which poses are equivalent or ambiguous
OCAM	Completed object + CAD + symmetry	$t,r,s$	Extract Siamese 3D-CNN features and regress 9DoF with three MLP head groups

CAPCM is supervised by S2C-Completion and the authors’ SN-Indoor dataset. OCAM uses learned relative-pose regression rather than explicit ICP between its object / CAD branches. My analysis: Visually plausible completion must also support final alignment. Strong ablations show task-specific benefits, but separate training leaves a risk of objective mismatch.

5. Experimental analysis

Setting	Class avg	Weighted avg	Interpretation
CAOA, no completion	48.63	60.92	Raw partial point clouds substantially limit alignment
+ CAPCM, no SEM	67.45	75.83	Completion provides the main gain
+ CAPCM + SEM, predicted segmentation	69.17	77.51	Symmetry adds approximately 2%
+ CAPCM + SEM, GT segmentation	78.15	86.54	Frontend segmentation still causes approximately 10% loss

On completion benchmarks, ODGNet trained with SN-Indoor + S2C-Completion + 100 cm context reaches CD-L1 / CD-L2 of 22.471 / 2.3 ( $\times10^{-3}$ ), better than training on generic ShapeNet or PCN data alone. Reported runtime is approximately 10.3 ms/object for completion, 3.5 ms/object for pose, and 0.58 s per end-to-end scene; segmentation must still be included in runtime accounting.

6. Robustness, strengths, limitations, and future directions

Occlusion / noise: CAPCM targets noisy incomplete scans and is evaluated on real Scan2CAD data, but cannot fully compensate for incorrect frontend segmentation.

Symmetry: SEM learns symmetry features in canonical CAD coordinates, providing explicit object-level priors. Generalization to approximately symmetric shapes outside its symmetry labels remains insufficiently quantified.

Strengths: A practical semantic-reconstruction setting, explicit 9DoF scale, and connected data, completion, alignment, runtime, and ablation evidence.

Limitations: CAD retrieval is outside the core method; completion and alignment heads are trained separately; a real end-to-end pipeline must still handle RGB-D depth bias, incorrect detections, and multiple similar CAD models.

Possible improvements: Backpropagate pose residuals into completion and use alignment uncertainty to decide whether to complete / abstain. Fit multiple CAD-retrieval hypotheses instead of assuming one correct model.

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.