Vojtěch Kloda

Searching in highly homogeneous marine video collections remains a challenging task even in the era of large multimodal deep networks.This difficulty arises because marine video content often falls outside the everyday vocabulary of users, while the visual homogeneity leads to frequent abstract labels such as fish or coral.This paper investigates the effectiveness of supplementing text queries with spatial location information to better address specific search intents.To effectively process location information, we investigate several strategies for video frame partitioning and content representation, formally defining and experimentally testing both static grid and dynamic segmentation variants.Furthermore, we present the results of a large-scale cognitive user study involving 220 participants.The study utilizes standard working memory interference techniques, comprising a target exposure phase, a distraction period, and a subsequent query specification from memory.The user queries are then evaluated against the proposed spatial retrieval models to determine how human spatial memory decay impacts retrieval effectiveness.Our results reveal that while dynamic segmentation models achieve the highest theoretical retrieval bounds under perfect conditions, their strict geometric boundaries are severely affected by human memory decay.Consequently, static overlapping grids demonstrate high robustness, outperforming dynamic models under more realistic, memory-driven search constraints.Finally, our experiments show that the observed spatial annotation perturbations can be modeled using a four-dimensional Kernel Density Estimation (KDE) method, enabling the simulation of realistic human memory decay on ideal bounding boxes.