ISeeCube is a visual analytics system for exploring eye-tracking data recorded from video stimuli. It combines spatiotemporal analysis in a space-time cube (STC) with AOI-based timeline visualizations and interactive scanpath comparison to support a wide range of analysis tasks.
Input. The system takes a video stimulus and gaze data from multiple viewers. Gaze data is represented as per-frame x/y coordinates for each viewer (pre-filtered fixations or raw data).
Preprocessing. Three automated steps prepare the data for exploration: (1) gaze density is computed as the distance of each point to the frame centroid of all viewers; (2) shot boundaries are detected in the video; (3) spatiotemporal clustering via mean shift (within shots) identifies groups of co-located gaze points indicating potential areas of interest.
STC visualization. Gaze data is arranged in a 3D space-time cube (two spatial axes, one temporal axis). Points are coloured by density, clusters shown as smoothed hulls, and wall projections provide 2D summaries. The analyst identifies timespans of attentional synchrony (dense single-cluster regions) and distributed attention (multiple clusters), plus motion signatures on projection walls.
AOI annotation. Informed by cluster positions, the analyst defines dynamic areas of interest using an editor that supports axis-aligned bounding boxes with key-frame interpolation. AOIs can be categorised and imported from ViPER-GT format.
AOI-based analysis. Gaze points are mapped to AOIs frame-by-frame (point-in-polygon, nearest-centre disambiguation). This yields per-viewer scanpath strings encoding the temporal sequence of fixated AOIs. A coordinated timeline visualization shows: (a) an AOI timeline with colour-coded existence bars and attention histograms; (b) viewer scarf plots showing frame-wise AOI assignment colour-coded by AOI identity; (c) AOI overview panels with filmstrips and histograms of attention, size, and position.
Scanpath clustering. For a selected timespan, pairwise viewer similarity is computed using one of three measures: Levenshtein distance (temporal+sequential coherence), attention map difference (overall dwell-time distribution), or transition map difference (AOI-to-AOI transition frequencies). Hierarchical agglomerative clustering with average linkage groups similar viewers, visualised as a dendrogram alongside reordered scarf plots.
Exploration loop. The analyst iteratively adjusts parameters (kernel size for density filtering, cluster size threshold), selects timespans, chooses similarity measures, shows/hides AOI categories, and interprets resulting patterns — identifying attentional synchrony, motion signatures, attention distribution, outlier viewers, and multiple viewer groups. The cycle continues until research questions about viewer behaviour are sufficiently answered.