跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-10-14AI 評分35

RTAformer:用於影片場景圖生成的關係轉變感知 Transformer

RTAformer: Relation-Transition-Aware Transformer for Video Scene Graph Generation

AI 導讀

RTAformer 是一種用於影片場景圖生成(VidSGG)的 Transformer,獲 ACCV 2026 接收,旨在解決關係改變後,時間模型仍保留過時關係預測的問題。

正文

We introduce RTAformer, accepted at ACCV 2026, for Video Scene Graph Generation (VidSGG).

Existing temporal models often retain outdated relation predictions even after the underlying relationships change. We identify this failure mode as Relational Semantic Inertia.

RTAformer addresses this issue by disentangling entity persistence from relation dynamics, using calibrated relation-query differences to guide predicate updates and cross-frame instance alignment to stabilize temporal representations.

Experiments on Action Genome demonstrate improved performance, with particularly strong gains on relation-transition frames under matched-backbone comparisons.

Paper: https://arxiv.org/abs/2610.20091

Code (coming soon): https://github.com/sejong-rcv/RTAformer

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co