Research map / Detection and segmentation / Open-set and few-shot detection
Grounded 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 Conditional Slot Attention (CoSA), which replaces slot attention's single random initialisation distribution with a learned Grounded Slot Dictionary of codebook vectors and per-object Gaussian slot priors, enabling object-type-specialised slots, dynamic slot-count estimation, and few-shot task adaptation.
Problem
Slot Attention initialises all slots from a single shared distribution, so it cannot learn specialised slots that bind to specific object types or stay invariant to identity-preserving appearance changes.
Approach
A Grounded Slot Dictionary pairs vector-quantised canonical object codebook entries with learnable Gaussian priors; scene features are abstracted via spectral decomposition to pick codebook entries, and slots are sampled from the matching per-object Gaussians before iterative slot attention refinement.
Key results
- On object discovery, CoSA reaches foreground ARI 0.96 on CLEVR and 0.83 on ObjectsRoom versus 0.93 and 0.78 for SA, and 0.32 on COCO versus 0.28 for DINOSAUR.
- On FMNIST2 addition-to-subtraction few-shot adaptation (k=100), CoSA attains 60.24% accuracy versus 11.06% for SA and 9.71% for Block-Slot.
- Dynamic slot-number estimation reduces FLOPs by up to 38% on CLEVR, with slot-count MAE of 2.06 on CLEVR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoSA with a grounded slot dictionary instead of BlockSlot | object discovery · CLEVR | foreground ARI: 0.94 ± 0.001 -> 0.96 ± 0.002 | Helped |
| Use CoSA with a grounded slot dictionary instead of SlotVAE | object discovery · ObjectsRoom | foreground ARI: 0.79 ± 0.01 -> 0.83 ± 0.002 | Helped |
| Use CoSA with a grounded slot dictionary instead of DINOSAUR | object discovery · COCO | foreground ARI: 0.28 ± 0.02 -> 0.32 ± 0.01 | Helped |
| Use CoSA instead of Block-Slot | object discovery · CLEVR | MSE: 3.73 -> 3.14 | Helped |
| Use CoSA instead of Block-Slot | object discovery · CLEVR | FID: 34.11 -> 29.12 | Helped |
| Use CoSA instead of IMPLICIT | object discovery · Tetrominoes | FID: 0.41 -> 0.41 | No clear effect |
| Use CoSA instead of Block-Slot | object discovery · Objects-Room | FID: 36.33 -> 28.19 | Helped |
| Use CoSA instead of IMPLICIT | object discovery · FFHQ | MSE: 47.06 -> 33.37 | Helped |
| Use CoSA instead of SA | few-shot reasoning adaptation · FMNIST2-Sub target | accuracy: 11.06 -> 60.24 | Helped |
| Use CoSA instead of CNN | few-shot reasoning adaptation · FMNIST2-Add target | accuracy: 12.35 -> 63.29 | Helped |
Surprising findings
- Standard CNN and vanilla SA classifiers perform near chance (around 10%) on cross-task transfer while CoSA transfers to 60%.
Limitations
- Limited variation in slot prompting and composition
- Assumes no representation overcrowding
- Assumes a maximum object area (eigenvalue-based slot estimation)
Experimental setup
Datasets: CLEVR Tetrominoes Objects-Room Bitmoji FFHQ COCO CLEVR-Hans3 CLEVR-Hans7 FloatingMNIST-2 FloatingMNIST-3 MNIST
Benchmarks: CLEVR Tetrominoes Objects-Room COCO Bitmoji FFHQ CLEVR-Hans3 CLEVR-Hans7 FloatingMNIST-2 FloatingMNIST-3
Metrics: ARI MSE FID SFID OPI accuracy F1 HMC MAE
Baselines: SA BlockSlot SlotVAE DINOSAUR IMPLICIT CNN Block-Slot
Builds on: Slot Attention VQ-VAE IODINE Gumbel-softmax
Closest papers by meaning
- Adaptive Slot Attention: Object Discovery with Dynamic Slot Number (CVPR 2024 · similarity 0.92)
- Identifiable Object-Centric Representation Learning via Probabilistic Slot Attention (NeurIPS 2024 · similarity 0.91)
- MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning (NeurIPS 2025 · similarity 0.91)
- GLASS: Guided Latent Slot Diffusion for Object-Centric Learning (CVPR 2025 · similarity 0.91)
- Slot-Guided Adaptation of Pre-trained Diffusion Models for Object-Centric Learning and Compositional Generation (ICLR 2025 · similarity 0.90)
- Cycle Consistency Driven Object Discovery (ICLR 2024 · 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.