Research map / Video / Video segmentation and tracking
Smoothing Slot Attention Iterations and Recurrences
Rongzhen Zhao, Wenyan Yang, Juho Kannala, Joni Pajarinen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add query preheater (Transformer decoder block) with preheat loss (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 56.9 -> 68.3 | Improved |
| use Transformer decoder block instead of Slot Attention module as preheater (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 63.3 -> 68.3 | Improved |
| switch cross-attention and self-attention order in preheater (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 49.6 -> 68.3 | Improved |
| stop gradient on preheated queries (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 67.5 -> 68.3 | Improved |
| preheating loss weight 100 instead of 10 (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 59.7 -> 68.3 | Improved |
| preheating loss weight 100 instead of 200 (ablation) | SmoothSA-i · COCO ablation | ARI + ARIfg: 67.4 -> 68.3 | Improved |
| share SA weights between first and non-first transforms instead of separate (ablation) | SmoothSA-v · YTVIS ablation | ARI + ARIfg: 52.3 -> 68.3 | Improved |
| use 3 first-frame iterations instead of 1 (unconditional video) (ablation) | SmoothSA-v · YTVIS unconditional video ablation | ARI + ARIfg: 97.4 -> 105.6 | Improved |
| use 1 non-first-frame iteration instead of 3 (unconditional video) (ablation) | SmoothSA-v · YTVIS unconditional video ablation | ARI + ARIfg: 103.4 -> 105.6 | Improved |
| use 3+1 iterations instead of 1+1 (conditional video) (ablation) | SmoothSA-v · MOVi-C conditional video ablation | ARI + ARIfg: 133.9 -> 136.3 | Improved |
| use 3+1 iterations instead of 3+3 (conditional video) (ablation) | SmoothSA-v · MOVi-C conditional video ablation | ARI + ARIfg: 132.7 -> 136.3 | Improved |
| replace DIAS SA with SmoothSA-i | DIAS · COCO object discovery | mIoU: 30.2 -> 31.8 | Improved |
Closest papers by meaning
- Adaptive Slot Attention: Object Discovery with Dynamic Slot Number (CVPR 2024 · similarity 0.88)
- MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning (NeurIPS 2025 · similarity 0.88)
- Grounded Object-Centric Learning (ICLR 2024 · similarity 0.87)
- SPOT: Self-Training with Patch-Order Permutation for Object-Centric Learning with Autoregressive Transformers (CVPR 2024 · similarity 0.86)
- Guided Slot Attention for Unsupervised Video Object Segmentation (CVPR 2024 · similarity 0.86)
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 2025 · similarity 0.86)
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.
