Get started
Work

Temporal Stride as an Inductive Bias in Masked Video Pretraining for Endoscopic Video Analysis

Year: 2026
Type: article
Abstract: Temporal stride is usually treated as a compute setting in masked video modeling, but this study shows that it is actually a major design choice that selects between different useful biases. In a controlled study with VideoMAE v2 on two clinical video datasets, REAL-Colon for diagnostic colonoscopy and GynSurg for laparoscopic surgery, it is observed that increasing stride shifts the encoder from focusing on high-frequency, texture-based details to focusing on lower-frequency, shape-based structure. This change is consistently observed across multiple analyses, including downstream clinical separability, attention patterns, manifold structure, spectral bias, token behavior, and temporal dynamics. Probing performance improves for polyp detection from F $1=0.659$ to 0.713, and for surgical action recognition from F $1=0.411$ to 0.507. Across both datasets these improvements follow a consistent monotonic trend across temporal strides, with Kendall rank correlations between stride and the primary metrics reaching + 1.00. At the same time, sparse pretraining makes the model less sensitive to fine temporal continuity, as shown by the Arrow of Time and Out-of-Order Frame Localization tasks. An information-theoretic argument is also provided that this stride effect comes from the input data distribution, not from the reconstruction target, which suggests that the same effect should apply more broadly across masked video modeling methods. The practical recommendation is to choose stride based on the temporal scale of the task: sparse strides (16-30) for pathology detection and surgical phase recognition, and dense stride (1-8) for tasks that depend on temporal causality.
Source: IEEE Access
DOI: 10.1109/access.2026.3725803
Language: English
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7203829038
FWCI:
Cites: 63
Cited by:
Related to: 10