Paper Overview
- Research field: Computer Vision
- Authors: Jiwook Han, Geo Ahn, Youngrae Kim, Jinwoo Choi
- Published: 2026-03-26
- arXiv: 2603.25733v1
- SlotVTG introduces a lightweight slot adapter that applies slot attention to decompose visual tokens into abstract slots and then reconstructs the original sequence.
- Object priors from self-supervised visual models encourage the formation of semantically coherent slots.
- The adapter approach avoids costly full multi-stage retraining of the MLLM.
- Significantly improves OOD robustness compared to standard fine-tuned MLLMs.
- Maintains competitive in-domain (ID) performance.
- Adds only minimal computational overhead.
Summary
Multimodal Large Language Models (MLLMs) have shown strong performance on Video Temporal Grounding (VTG). However, their coarse recognition capabilities are insufficient for fine-grained temporal understanding, making task-specific fine-tuning indispensable. This fine-tuning causes models to memorize dataset-specific shortcuts rather than faithfully grounding in the actual visual content, leading to poor Out-of-Domain (OOD) generalization.
Object-centric learning offers a promising remedy by decomposing scenes into entity-level representations, but existing approaches require re-running the entire multi-stage training pipeline from scratch. The authors propose SlotVTG, a framework that steers MLLMs toward object-centric, input-grounded visual reasoning at minimal cost.
Method
Results
Cross-domain evaluations on standard VTG benchmarks show that SlotVTG:
*Auto-collected on 2026-03-28*