Research map / Detection and segmentation / Open-set and few-shot detection
Hierarchical Compact Clustering Attention (COCA) for Unsupervised Object-Centric Learning
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 COCA-Net, a bottom-up hierarchical encoder of compact clustering attention layers plus a Spatial Broadcast Decoder, for unsupervised object discovery on single images, evaluated on six synthetic datasets.
Problem
Slot Attention-based unsupervised object-centric learning relies on K-Means-like soft clustering, requiring a predefined number of slots and struggling with background segments and initialization sensitivity on synthetic and real images.
Approach
A COCA layer performs attention-based hierarchical agglomerative clustering within non-overlapping windows, scoring candidate affinity masks by moment-of-inertia-based compactness and sequentially extracting the most compact mask as a cluster (SBC-style concealing), with cascaded layers forming COCA-Net and dynamic cluster numbers.
Key results
- On ObjectsRoom, COCA-Net encoder-side foreground ARI 0.87 vs BOQ-SA 0.59 and INV-SA 0.80 (Table 2).
- On ShapeStacks, encoder foreground mSC improves nearly 30 points over baselines, e.g., 0.85 vs 0.59 for BOQ-SA.
- Ablation: compactness-based anchor selection raises ObjectsRoom decoder ARI from 0.832 (random anchor, COCA-Net-RAS) to 0.894.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace compactness-based anchor selection with random anchor-node selection (ablation) | COCA-Net · ObjectsRoom, decoder foreground | ARI: 0.894 -> 0.832 | Hurt |
| replace compactness-based anchor selection with random anchor-node selection (ablation) | COCA-Net · ObjectsRoom, decoder foreground | mSC: 0.832 -> 0.739 | Hurt |
| use dynamic output-slot generation with a stopping condition (ablation) | COCA-Net · CLEVR10, decoder foreground | ARI | Helped |
| use dynamic output-slot generation with a stopping condition (ablation) | COCA-Net · CLEVR10, decoder foreground | mSC | Helped |
Surprising findings
- COCA-Net's background-inclusive ARI on ShapeStacks is poor (0.31), which the authors attribute to ShapeStacks providing a single mask for all background segments.
Limitations
- Bottom-up clustering lacks top-down feedback, so early-layer errors cannot be corrected.
- Moment-of-inertia reference circle computation creates heavy memory requirements from pairwise density comparisons.
- Background-inclusive ARI is poor on ShapeStacks due to single background mask annotation.
Experimental setup
Datasets: Tetrominoes Multi-dSprites ObjectsRoom ShapeStacks CLEVR6 CLEVRTex CLEVR10
Benchmarks: Tetrominoes Multi-dSprites ObjectsRoom ShapeStacks CLEVR6 CLEVRTex
Metrics: ARI mSC MSE
Baselines: GENESIS-v2 INV-SA BOQ-SA
Builds on: Slot Attention GENESIS-v2 Spatial Broadcast Decoder ViT-22B Stick-Breaking Clustering
Closest papers by meaning
- Finding Distributed Object-Centric Properties in Self-Supervised Transformers (CVPR 2026 · similarity 0.86)
- unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning (ICML 2025 · similarity 0.84)
- Training-Free Open-Ended Object Detection and Segmentation via Attention as Prompts (NeurIPS 2024 · similarity 0.84)
- Attention-based clustering (NeurIPS 2025 · similarity 0.84)
- Bootstrapping Top-down Information for Self-modulating Slot Attention (NeurIPS 2024 · similarity 0.83)
- HEAP: Unsupervised Object Discovery and Localization with Contrastive Grouping (AAAI 2024 · similarity 0.83)
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.