Research map / Detection and segmentation / Open-set and few-shot detection
MetaSlot: Break Through the Fixed Number of Slots in 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 MetaSlot, a plug-and-play Slot Attention variant that uses a VQ prototype codebook to prune duplicate slots in a two-stage aggregation and injects annealed Gaussian noise, so the number of object slots adapts to the scene.
Problem
Standard Slot Attention uses a fixed number of slots and random initialization, so objects get split across multiple slots when object counts vary, hurting object discovery and representation quality.
Approach
A global VQ codebook of object prototypes quantizes first-stage slots; duplicate slots mapping to the same prototype are masked out, and the surviving prototype slots initialize a second, masked Slot Attention stage with linearly annealed Gaussian noise injected into features, with the codebook updated by mini-batch k-means EMA.
Key results
- On COCO with DINOSAUR decoder (7 slots), MetaSlot raises FG-ARI from 35.0 to 40.3 and mBO from 28.3 to 29.5 over the BO-QSA-based baseline.
- On ClevrTex, MetaSlot-Mlp improves ARI from 50.7 to 64.6 over DINOSAUR.
- On YTVIS(HQ) with VideoSAUR, MetaSlot improves FG-ARI by +18.3 and mBO by +2.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the baseline aggregator with MetaSlot | DINOSAUR · COCO | FG-ARI: 35.0 ± 1.2 -> 40.3 ± 0.5 | Helped |
| replace the baseline aggregator with MetaSlot | DINOSAUR · VOC | ARI: 21.5 ± 0.7 -> 32.2 ± 1.2 | Helped |
| replace the baseline aggregator with MetaSlot | VideoSAUR · YTVIS(HQ) | FG-ARI: 49.0 ± 0.9 -> 67.3 ± 2.1 (+18.3) | Helped |
| replace the DINOSAUR aggregator with MetaSlot | COCO set prediction model · COCO | top1 accuracy: 0.33 -> 0.36 | Helped |
| set codebook size to 1024 instead of 512 (ablation) | MetaSlot · COCO | FG-ARI: 40.3 -> 40.4 | No clear effect |
| set slot number to 15 instead of 7 (ablation) | MetaSlot · COCO | FG-ARI: 40.3 -> 36.0 | Hurt |
| remove prototype-guided initialization (ablation) | MetaSlot · COCO | FG-ARI: 40.2 -> 40.3 | Helped |
| disable prototype-based pruning (ablation) | MetaSlot · COCO | FG-ARI: 40.0 -> 40.3 | Helped |
| increase BO-QSA iteration count to 6 (ablation) | BO-QSA · COCO | FG-ARI: 35.0 ± 1.2 -> 37.9 | Helped |
| remove progressively attenuated noise during slot updates (ablation) | MetaSlot_Mlp · COCO | ARI | Mixed |
| remove progressively attenuated noise during slot updates (ablation) | MetaSlot_Mlp · COCO | mBO | Helped |
Surprising findings
- Increasing BO-QSA iterations to 6 does not improve its performance (FG-ARI 37.9 vs 35.0 at default is not a gain over baseline in the same way; the authors note iteration count alone does not help).
- MetaSlot reports comparable or better results than SPOT even without SPOT's two-stage self-distillation training.
Limitations
- Absolute positional encoding makes the model non-equivariant to translations, so prototypes may encode position-dependent patterns.
- Prototypes mainly capture global shape or semantics, missing finer attribute compositions.
- Not evaluated under out-of-distribution conditions.
Experimental setup
Datasets: ClevrTex MS COCO 2017 PASCAL VOC 2012 YTVIS(HQ) MOVi-C
Benchmarks: ClevrTex MS COCO 2017 PASCAL VOC 2012 YTVIS(HQ) MOVi-C
Metrics: ARI FG-ARI mBO mIoU top-1 accuracy R^2
Baselines: Slot Attention BO-QSA DINOSAUR SLATE SlotDiffusion SPOT VideoSAUR AdaSlot SlotContrast SysBinder NLoTM
Builds on: Slot Attention DINOSAUR BO-QSA SLATE SlotDiffusion
Closest papers by meaning
- Adaptive Slot Attention: Object Discovery with Dynamic Slot Number (CVPR 2024 · similarity 0.93)
- Grounded Object-Centric Learning (ICLR 2024 · similarity 0.91)
- GLASS: Guided Latent Slot Diffusion for Object-Centric Learning (CVPR 2025 · similarity 0.89)
- SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers (CVPR 2024 · similarity 0.88)
- Identifiable Object-Centric Representation Learning via Probabilistic Slot Attention (NeurIPS 2024 · similarity 0.88)
- Slot-Guided Adaptation of Pre-trained Diffusion Models for Object-Centric Learning and Compositional Generation (ICLR 2025 · similarity 0.88)
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.