🎯 Object Detection¶

🧪 ICML2026 · 1 paper notes

Smoothing Slot Attention Iterations and Recurrences: Addressing two long-standing but overlooked issues in Slot Attention for image and video object-centric learning—namely, "insufficient information in cold-start queries" and "forced unification of aggregation transformations for first/non-first frames"—the authors propose SmoothSA: a self-distillation-based lightweight warm-up module that injects sample-specific information into queries, and a scheduling scheme where the first frame undergoes three full iterations while non-first frames only run one. This approach achieves new SOTA on both image and video OCL benchmarks.