Peak-End-Net: A Peak-End Rule Inspired Framework for Generalizable Video Aesthetic Assessment

TL;DR — Peak-End-Net is a lightweight, interpretable framework for video aesthetic assessment (VAA) that borrows the psychological "peak-end rule" to weight salient moments and endings, achieving state-of-the-art results on two benchmarks.

• Transfers knowledge from image aesthetic assessment (IAA) via a pretrained IAA head that produces frame-wise aesthetic priors, used to identify salient moments and guide peak-end-based temporal aggregation.
• Adds an aesthetic rhythm encoder to model how a video evolves aesthetically over time, plus a dynamic gated fusion mechanism for robustness under distribution shift.
• Built on a frozen ViT with only a small number of trainable parameters, making it parameter-efficient and scalable.
• Reports state-of-the-art performance on in-domain VADB and cross-domain DIVIDE-3K; code and models released on GitHub.