pipeline_low_power.yaml

# MindSight -- Low-Power Pipeline Configuration
#
# UNVALIDATED throughput profile derived from the known-good preset
# (configs/pipeline_known_good.yaml, KG_Standard 2026-07-09) for weaker
# hardware (no discrete GPU, older laptops). Same study semantics -- all
# phenomena on, Gaze-LLE blend on -- with lighter models and cheaper
# detection. A/B against the known-good preset on your own footage before
# trusting numbers for analysis; record validated values in KNOWN_GOOD.md.
#
# Differences from the known-good preset, all throughput-motivated:
#   mgaze_model    mobileone_s0 (lightest gaze backend; device-switching)
#   detect_scale   0.75 (detection on a 3/4-size frame)
#   fast           true (skip phenomena on non-detection frames, throttled
#                  dashboard bridge and GUI poll rate)
#   min_call_gap   40 (fewer Gaze-LLE inferences)
#
# Usage:
#   python MindSight.py --pipeline configs/pipeline_low_power.yaml \
#       --source path/to/video.mp4

# ── Object detection ─────────────────────────────────────────────────────────
detection:
  model: "yolov8n.pt"
  conf: 0.25
  detect_scale: 0.75
  vp_model: "yoloe-26l-seg.pt"
  skip_frames: 1
  obj_persistence: 0

# ── Gaze / ray geometry ──────────────────────────────────────────────────────
gaze:
  ray_length: 1.3
  adaptive_ray: "off"
  snap_dist: 150.0
  snap_bbox_scale: 0.0
  snap_w_dist: 1.0
  snap_w_angle: 0.8
  snap_w_size: 0.0
  snap_w_intersect: 0.5
  snap_w_temporal: 0.3
  snap_gate_angle: 60.0
  snap_head_blend: 0.3
  snap_quality_thresh: 0.8
  snap_tip_dist: -1.0
  snap_tip_quality: -1.0
  conf_ray: false
  gaze_tips: true
  tip_radius: 70
  gaze_cone: 5.0
  gaze_lock: false
  dwell_frames: 15
  lock_dist: 100
  gaze_debug: false
  snap_release_frames: 5
  snap_engage_frames: 0
  reid_grace_seconds: 4.5
  hit_conf_gate: 0.0
  detect_extend: 0.0
  detect_extend_scope: both

# ── Output ───────────────────────────────────────────────────────────────────
output:
  anonymize_padding: 0.3

# ── Performance ──────────────────────────────────────────────────────────────
performance:
  fast: true
  skip_phenomena: 0
  lite_overlay: false
  no_dashboard: true
  profile: false

# ── Depth (disabled) ─────────────────────────────────────────────────────────
depth:
  enabled: false
  backend: midas_small
  input_size: 384
  skip_frames: 1
  depth_aware_scoring: false
  snap_w_depth: 0.4

# ── Phenomena trackers (all on, same as known-good) ──────────────────────────
phenomena:
  - joint_attention:
      ja_window: 0
      ja_quorum: 1.0
      ja_window_thresh: 0.7
  - mutual_gaze
  - social_referencing
  - gaze_following
  - gaze_aversion
  - scanpath
  - gaze_leadership
  - attention_span

# ── Plugin / backend settings ────────────────────────────────────────────────
plugins:
  all_phenomena: true

  # Lightest MobileGaze build; extensionless = device-switching
  # (mobileone_s0.pt on CUDA, mobileone_s0_gaze.onnx elsewhere).
  mgaze_model: "mobileone_s0"
  mgaze_dataset: "gaze360"

  # Gaze-LLE blend stays ON (primary operating mode), lighter cadence.
  # W4A: the 16 MB DINOv3-distilled pico ONNX default suits low power --
  # no torch-hub backbone download, CPU-only, more accurate than the
  # torch ViT-B/14 engine it replaces.
  rf_gazelle_model: "gazelle_hgnetv2_pico_inout_distill_1x3x640x640_1xNx4.onnx"
  rf_gazelle_name: "gazelle_dinov2_vitb14"
  gazelle_name: "gazelle_dinov2_vitb14"
  min_call_gap: 40
  gazelle_device: auto
  gazelle_inout_threshold: 0.5
  gazelle_skip_frames: 0
  gazelle_fp16: false
  gazelle_compile: false
  dir_beta: 0.5
  len_beta: 0.3
  len_hold_tau: 5.0
  fixation_v_threshold: 0.04
  fixation_d_threshold: 0.15
  dir_min_cutoff: 1.0
  len_min_cutoff: 1.0

  forward_gaze_threshold: 13.0
  smooth_snap: all
  smooth_snap_alpha: 0.9
  gaze_leader_tips: false
  gaze_leader_tip_lag: 15

  merge_overlaps: true
  merge_overlap_strategy: "dynamic"
  merge_overlap_threshold: 0.55