Research map / Detection and segmentation / Open-set and few-shot detection
Adaptive Slot Attention: Object Discovery with Dynamic Slot Number
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 AdaSlot, an object-centric auto-encoder that dynamically selects a subset of slots per image via Gumbel-Softmax mean-field sampling and a masked slot decoder, removing the need to predefine slot count.
Problem
Slot attention models require a predefined number of slots, which needs dataset knowledge and ignores per-instance variation in object count, hurting object discovery quality.
Approach
Generate Kmax slots, then learn a per-slot keep/drop decision with Gumbel-Softmax straight-through sampling (mean-field factorization), decode only kept slots via a zero-mask mixture decoder, and regularize the expected number of kept slots to favor fewer slots.
Key results
- On MOVi-C, AdaSlot reaches FG-ARI 75.59 versus 73.23 for the best fixed-slot DINOSAUR (6 slots).
- On MOVi-E, AdaSlot reaches FG-ARI 76.73 and F1 81.42, the best among compared models (best fixed DINOSAUR 9 slots: ARI 76.01).
- On MS COCO, AdaSlot reaches ARI 39.00, roughly 20 points above the 33-slot DINOSAUR, but slightly below the best fixed-slot setting (6 slots, ARI 41.89).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MOVi-C | ARI: 73.23 -> 75.59 | Helped |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MOVi-C | F1: 82.56 -> 84.25 | Helped |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MOVi-E | ARI: 76.01 -> 76.73 | Helped |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MOVi-E | mBO: 30.54 -> 29.83 | Hurt |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MS COCO 2017 | ARI: 41.89 -> 39.00 | Hurt |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MS COCO 2017 | mBO: 27.77 -> 27.36 | Hurt |
| Use adaptive slot selection instead of fixed slot counts | DINOSAUR · MS COCO 2017 | ARI: 20.83 -> 39.00; 41.89 | Mixed |
| Use zero-slot masking instead of zero-mask decoding (ablation) | AdaSlot masked slot decoder · Reconstruction and object grouping | reconstruction quality | Hurt |
| Use learnable-slot masking instead of zero-mask decoding (ablation) | AdaSlot masked slot decoder · Reconstruction and object grouping | reconstruction quality | Hurt |
Surprising findings
- On COCO, the best fixed-slot DINOSAUR (6 slots) marginally outperforms AdaSlot, which the authors attribute to incomplete labels and cluttered scenes.
Limitations
- Struggles on complex densely packed scenes like COCO with incomplete annotations.
- Performance on small, dense objects is limited.
- Does not address real-world part-whole hierarchies.
Experimental setup
Datasets: CLEVR10 MOVi-C MOVi-E MS COCO 2017
Benchmarks: CLEVR10 MOVi-C MOVi-E MS COCO 2017
Metrics: FG-ARI precision recall F1 mBO CorLoc Purity AMI NMI Jaccard
Baselines: DINOSAUR GENESIS-V2
Builds on: Slot Attention DINOSAUR Gumbel-Softmax mean-field variational inference
Closest papers by meaning
- MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning (NeurIPS 2025 · similarity 0.93)
- Grounded Object-Centric Learning (ICLR 2024 · similarity 0.92)
- Identifiable Object-Centric Representation Learning via Probabilistic Slot Attention (NeurIPS 2024 · similarity 0.91)
- GLASS: Guided Latent Slot Diffusion for Object-Centric Learning (CVPR 2025 · similarity 0.90)
- Cycle Consistency Driven Object Discovery (ICLR 2024 · similarity 0.90)
- Slot-Guided Adaptation of Pre-trained Diffusion Models for Object-Centric Learning and Compositional Generation (ICLR 2025 · similarity 0.89)
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.