Research map / Detection and segmentation / Open-set and few-shot detection
unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning
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 unMORE, a two-stage unsupervised multi-object segmentation pipeline that learns existence, center-field and boundary-distance objectness on ImageNet, then uses a network-free center-boundary reasoning module to discover multiple objects in scene images.
Problem
Unsupervised multi-object segmentation on single images: existing reconstruction-based slot methods work mainly on synthetic data, while pretrained-feature grouping methods under-segment and discover too few objects in real-world scenes.
Approach
Train an objectness network on ImageNet to predict three object-centric fields (binary existence, unit-vector center field, normalized signed boundary distance field) from pseudo masks, then iteratively refine bounding boxes with a network-free reasoning procedure that splits proposals at anti-center peaks and updates borders using boundary-distance gradients.
Key results
- On COCO* val, unMORE achieves 18.0 AP box and 15.5 AP mask versus 15.5 AP box for CuVLER, the strongest prior detector-based method.
- The network-free unMORE_disc reaches 10.1 AP box on COCO* without training a detector, versus 5.5 for VoteCut.
- On images with 15+ GT objects, unMORE_disc gets 7.1 AP box versus 1.8 for VoteCut, while baselines collapse on crowded images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use network-free center-boundary reasoning with the objectness network instead of VoteCut | COCO* direct object discovery · COCO* val | APbox: 5.5 -> 10.1 | Helped |
| train the detector with unMORE pseudo labels instead of CuVLER pseudo labels | COCO* detector-based segmentation · COCO* val | APbox: 15.5 -> 18.0 | Helped |
| use unMORE rather than the strongest listed baseline | zero-shot detection · COCO20K | APbox: 13.1 -> 13.9 | Helped |
| use unMORE rather than the strongest listed baseline | zero-shot detection · LVIS | APbox: 4.7 -> 5.6 | Helped |
| use unMORE rather than the strongest listed baseline | zero-shot detection · KITTI | APbox: 9.5 -> 13.7 | Helped |
| use unMORE rather than the strongest listed baseline | zero-shot detection · VOC | APbox: 21.5 -> 22.7 | Helped |
| use unMORE rather than the strongest listed baseline | zero-shot detection · Object365 | APbox: 11.5 -> 12.4 | Helped |
| use unMORE rather than CuVLER | zero-shot detection · OpenImages | APbox: 11.4 -> 11.2 | Hurt |
| add object existence scoring to center + boundary fields (ablation) | unMORE objectness representations · COCO* val | APbox: 17.0 -> 18.0 | Helped |
| add the object center field to existence + boundary distance field (ablation) | unMORE objectness representations · COCO* val | APbox: 17.1 -> 18.0 | Helped |
| add the boundary distance field to existence + center field (ablation) | unMORE objectness representations · COCO* val | APbox: 15.4 -> 18.0 | Helped |
| add the object existence score to center field only (ablation) | unMORE objectness representations · COCO* val | APbox: 15.8 -> 14.2 | Hurt |
Surprising findings
- The learned boundary distance field alone gives the largest ablation gain, outperforming the combined center and existence fields in several settings.
- UnSAM has high AR scores when unlimited predictions are allowed but low AP and AR at 100 predictions, indicating over-segmentation.
Limitations
- Performance is likely limited by the learned objectness priors from natural-image training data, with weak generalization to domains such as medical images.
- The reasoning module relies on hand-set thresholds (tau_e, tau_c, tau_b) and an adjustment ratio.
Experimental setup
Datasets: ImageNet COCO COCO20K LVIS VOC KITTI Object365 OpenImages
Benchmarks: COCO* COCO20K LVIS VOC KITTI Object365 OpenImages
Metrics: AP box AP mask AR box AR mask
Baselines: FreeMask MaskCut VoteCut DINOSAUR FOUND UnSAM CutLER CuVLER
Builds on: CuVLER CutLER DINO/v2 DPT
Closest papers by meaning
- Self-Supervised Learning of Object Parts for Semantic Segmentation (CVPR 2022 · similarity 0.90)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.89)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.89)
- Segment Anything without Supervision (NeurIPS 2024 · similarity 0.89)
- Unsupervised Universal Image Segmentation (CVPR 2024 · similarity 0.87)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.