Motion-aligned Temporal Adapter (MATA)
A parameter-efficient method that adapts a frozen CLIP video backbone by specializing its features along time. It uses (1) learnable temporal positional embeddings to align temporal phase and (2) a local motion transition operator implemented with depth-wise convolution to model short motion changes between nearby frames.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
In few-shot fine-grained video action recognition, actions can be visually similar and differ mainly by the order/timing of motion. Aligning temporal phase and explicitly modeling local motion transitions helps the model focus on these timing differences while keeping the backbone frozen.
What changed
Instead of updating the entire CLIP backbone for few-shot learning, MATA adds a temporal adapter module: it replaces/augments standard temporal handling with learnable temporal positional embeddings for phase alignment and depth-wise convolution for local motion transition modeling.
Where it helps
- few-shot fine-grained video action recognition
- recognizing actions differentiated mainly by motion timing
- settings with limited trainable parameters
- robustness needs under frame dropping or temporal disruptions
- video-only adaptation on top of a frozen CLIP backbone
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The approach is described for few-shot fine-grained video action recognition using video modality alone; effectiveness outside this setting (e.g., other modalities or tasks) is not specified.
Sources
- MATA: Motion-aligned Temporal Adapter ↗
arXiv · primary · Checked 2026-11-10
