Research map / Detection and segmentation / Open-set and few-shot detection

Multi-Part Object Representations via Graph Structures and Co-Part Discovery

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Object-centric learning theory and video · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ECO-Net, which represents multi-part objects as explicit part graphs and groups recurring, spatially related parts into objects via a co-part discovery algorithm, evaluated on simulated, realistic, and real-world datasets plus new occlusion and out-of-distribution benchmarks.

Problem

Unsupervised object-centric representation methods fail to recognize multi-part objects under occlusion and out-of-distribution conditions because part-whole relations are only implicitly encoded through indirect training objectives.

Approach

Build an explicit graph of Felzenszwalb parts (nodes: centroid-relative boundary vectors; edges: centroid offsets between parts) and cluster neighboring parts that recur with similar relations across a batch into objects, stored in an object memory, instead of learning part-whole structure implicitly.

Key results

• ECO-Net reaches 100.0 ARI/mDice/mIoU on Tetrominoes and 97.7 ARI, 99.4 mDice, 98.8 mIoU on AbsScene, versus best baseline BO-QSA/Slot Attention around 94-99 ARI but much lower mDice/mIoU.
• On occluded benchmarks ECO-Net achieves 93.1 mDice on AbsScene-O (best baseline Felzenszwalb 70.4) and 90.3 mDice on GSO-O (best baseline MaskCut 85.8).
• On out-of-distribution AbsScene-C ECO-Net achieves 93.5 ARI and 97.3 mDice, versus best baseline Felzenszwalb 60.7 ARI and DINOSAUR 82.0 ARI.

Surprising findings

• Felzenszwalb's classical segmentation is a strong baseline, outperforming many learned object-centric methods on several datasets.
• Replacing the raw pixel features with pretrained DINO features degraded ECO-Net on most datasets (e.g., Tetrominoes mIoU 100 to 57.3).

Limitations

• Authors note real-world scenes are of higher visual complexity than the datasets evaluated, an open problem.
• The co-part discovery algorithm is not yet efficient enough for real-time use; the worst-case complexity can reach O(M^4) when segment degree is large.
• The similarity threshold epsilon is hand-set rather than learned.

Experimental setup

Datasets: Tetrominoes AbsScene GSO SKU-110K PASCAL VOC MS COCO

Benchmarks: Tetrominoes AbsScene GSO SKU-110K PASCAL VOC MS COCO AbsScene-O GSO-O AbsScene-C

Metrics: ARI mDice mIoU R^2

Baselines: SLIC Ncut Felzenszwalb Slot Attention GENESIS-V2 BO-QSA DINOSAUR MaskCut OC-Net

Builds on: Felzenszwalb graph-based segmentation Slot Attention DINO MaskCut DINOSAUR

Closest papers by meaning

• Multi-Object Representation Learning via Feature Connectivity and Object-Centric Regularization (NeurIPS 2023 · similarity 0.91)
• Unsupervised Part Discovery from Contrastive Reconstruction (NeurIPS 2021 · similarity 0.90)
• Deep Co-Occurrence Feature Learning for Visual Object Recognition (CVPR 2017 · similarity 0.89)
• unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning (ICML 2025 · similarity 0.89)
• Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction (NeurIPS 2025 · similarity 0.87)
• Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.87)