Sharingan uses a ViT-12 + conditional DPT decoder to predict where each person is looking. Up to 3 people share a single forward pass. Requires WebGPU.
Total time:
Total frames: —
Samples saved: —
| Column | Description |
|---|---|
frame | Frame index (0-based) |
timestamp | Time in seconds |
face_index | Person index in frame (0–2) |
face_x1/y1/x2/y2 | Face bounding box (px) |
gaze_x / gaze_y | Predicted gaze target (px) |
inout_prob | In-frame gaze probability (0–1) |
| Gaze vector | |
gaze_vx / gaze_vy | Unit direction vector (face center → gaze target) |
gaze_angle_deg | Gaze angle in degrees (atan2, 0° = right) |
gaze_dist_px | Pixel distance: face center → gaze target |
| Heatmap metrics | |
hm_peak | Peak softmax probability (higher = more certain) |
hm_entropy | Normalised entropy (0 = certain, 1 = uniform) |
hm_spread_px | Weighted spatial std-dev of heatmap (px, RMS) |
| Joint Visual Attention (JVA) | |
jva_partner_idx | Face index of best JVA partner (−1 if alone) |
jva_gaze_dist_px | Pixel distance between the two gaze targets |
jva_dir_sim | Cosine similarity of gaze direction vectors (−1 … +1) |
jva_hm_overlap | Bhattacharyya coefficient of heatmaps (0–1) |
jva_score | Combined JVA score: proximity + direction + overlap (0–1) |
tag | User-defined label from the recording bar |
Your camera was not detected. If you are using Google Chrome, please follow these instuctions to allow the site to use your webcam. If you are not using Google Chrome, please use Google Chrome, as we do not currently support other browsers!