Research map / Detection and segmentation / Open-set and few-shot detection
Cycle Consistency Driven Object Discovery
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.
Adds two cycle-consistency losses (Slot-Feature-Slot and Feature-Slot-Feature) on slot-feature graphs to existing slot-based object-centric models, improving object discovery on synthetic and real data and downstream Atari and Causal World RL.
Problem
Unsupervised object discovery with slot-based representations relies on architectural priors that are unreliable, and the downstream utility of object-centric representations has been underexplored.
Approach
Treat slots and encoder features as nodes of a directed graph and add two auxiliary cycle-consistency losses that enforce each object's features to map to one slot and each slot to a distinct feature subset; the losses are plug-in to any slot-attention-style method, with an EMA encoder supplying feature-similarity supervision.
Key results
- On ClevrTex, adding both cyclic losses to Slot Attention raises FG-ARI from 0.6240 to 0.7245 (reported in the original paper; the authors' own reproduction of SA was 0.5864).
- On Atari, the DT + SA + Cyclic model beats the baseline Decision Transformer on 10 of 13 games (e.g., Pong 14.8 vs 11.0, Breakout 110.2 vs 70.6).
- On MOVi-E, DINOSAUR + Cyclic improves FG-ARI from 65.1 to 69.7 and mBO from 33.5 to 37.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add both cycle consistency objectives to Slot-Attention | Slot-Attention · ObjectsRoom, ShapeStacks, and ClevrTex | FG-ARI: ObjectsRoom 0.7819 -> 0.8169; ShapeStacks 0.7738 -> 0.7838; ClevrTex 0.6240 -> 0.7245 | Helped |
| Add both cycle consistency objectives to DINOSAUR | DINOSAUR (ViT-B/8) · MOVi-C | FG-ARI: 68.9 -> 72.4 | Helped |
| Add both cycle consistency objectives to DINOSAUR | DINOSAUR (ViT-B/8) · MOVi-E | FG-ARI: 65.1 -> 69.7 | Helped |
| Add both cycle consistency objectives to MoTok | MoTok · MOVi-E | FG-ARI: 67.38 -> 72.48 | Helped |
| Add cycle consistency objectives to BO-Slate | BO-Slate · COCO | AP@05: 16.64 -> 18.96 | Helped |
| Add cycle consistency objectives to BO-Slate | BO-Slate · ScanNet | AP@05: 24.67 -> 29.20 | Helped |
| Add cycle consistency objectives to BO-Slate | BO-Slate · Stanford Dogs, Stanford Cars, CUB200 Birds, and Flowers | IoU: Dogs 0.7875 -> 0.8456; Cars 0.7686 -> 0.8447; Birds 0.6129 -> 0.6497; Flowers 0.7461 -> 0.7745 | Helped |
| Add both cycle consistency objectives to Slot-Attention | Slot-Attention · Atari games | mean return: Cyclic outperforms DT in 10 out of 13 games | Mixed |
| Apply cycle objectives to all slot-attention iterations instead of only the last iteration (ablation) | Slot-Attention + Cyclic · ClevrTex | FG-ARI: 0.5453 -> 0.7245 | Helped |
| Add an EMA encoder (ablation) | Slot-Attention + Cyclic · ShapeStacks | FG-ARI: 0.7028 -> 0.7838 | Helped |
| Compute the Feature-Slot-Feature loss only for i=j rather than all i,j pairs (ablation) | Slot-Attention + Cyclic · ObjectsRoom | FG-ARI: 0.7341 -> 0.8169 | Helped |
| Compute the Feature-Slot-Feature loss only for i=j rather than all i,j pairs (ablation) | Slot-Attention + Cyclic · ShapeStacks | FG-ARI: 0.7161 -> 0.7838 | Helped |
Surprising findings
- Applying the Feature-Slot-Feature loss only on diagonal pairs (i = j) works better than applying it to all pairs, which the authors attribute to difficulty of learning accurate pairwise similarities from scratch.
- Using only the Slot-Feature-Slot loss on ShapeStacks sharply reduced FG-ARI variance-prone results (0.5491 ± 0.4521), and high lambda_sfs' causes performance degradation, possibly due to trivial single-slot solutions.
Limitations
- The authors could not reproduce the Slot Attention ClevrTex baseline result and reported the original paper's number instead.
- Results on the Shapestacks FG-ARI show high variance for some configurations, and the Causal World and Atari RL evaluations are limited to a few tasks.
- The authors note that downstream evaluation of object-centric representations on more complex, real-world tasks remains needed.
Experimental setup
Datasets: ObjectsRoom ShapeStacks ClevrTex MOVi-C MOVi-E COCO ScanNet Atari Causal World Stanford Dogs Stanford Cars CUB-200 Birds
Benchmarks: ObjectsRoom ShapeStacks ClevrTex MOVi-C MOVi-E COCO ScanNet Atari Causal World Stanford Dogs Stanford Cars CUB-200 Birds
Metrics: FG-ARI MSE mBO AP@05 PQ precision recall IoU Dice mean return
Baselines: Slot Attention SLATE DINOSAUR MoTok MONet IODINE Implicit Slot Attention BO-Slate Decision Transformer
Builds on: slot attention SLATE DINOSAUR decision transformer PPO
Closest papers by meaning
- Adaptive Slot Attention: Object Discovery with Dynamic Slot Number (CVPR 2024 · similarity 0.90)
- Grounded Object-Centric Learning (ICLR 2024 · similarity 0.89)
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 2025 · similarity 0.88)
- Identifiable Object-Centric Representation Learning via Probabilistic Slot Attention (NeurIPS 2024 · similarity 0.88)
- MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning (NeurIPS 2025 · similarity 0.87)
- GLASS: Guided Latent Slot Diffusion for Object-Centric Learning (CVPR 2025 · 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.