configs/ablations.yaml

# Ablation Study Configurations
#
# Each ablation isolates one variable to measure its contribution.
# Run: python scripts/run_ablations.py --config configs/ablations.yaml
#
# Baseline: SCRFD-34G with default settings

ablations:
  # 1. Sample Redistribution (SCRFD's key contribution)
  sample_redistribution:
    description: "Impact of large-scale crops on small face detection"
    experiments:
      - name: "baseline_crops"
        data.crop_scales: [0.3, 0.45, 0.6, 0.8, 1.0]
        note: "Standard crops only (no SR)"
      - name: "extended_crops"
        data.crop_scales: [0.3, 0.45, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0]
        note: "Full Sample Redistribution (SCRFD default)"
      - name: "extreme_crops"
        data.crop_scales: [0.3, 0.6, 1.0, 1.5, 2.0, 2.5, 3.0]
        note: "Even larger crops — test limits"

  # 2. Loss Function Choice
  loss_functions:
    description: "GFL vs Focal Loss for classification"
    experiments:
      - name: "gfl_diou"
        model.losses.cls: gfl
        model.losses.reg: diou
        note: "SCRFD default"
      - name: "focal_diou"
        model.losses.cls: focal
        model.losses.reg: diou
        note: "Standard focal loss"
      - name: "gfl_smoothl1"
        model.losses.cls: gfl
        model.losses.reg: smooth_l1
        note: "GFL + Smooth L1 regression"

  # 3. Anchor Matching Strategy
  matching:
    description: "ATSS vs fixed IoU threshold matching"
    experiments:
      - name: "atss_k9"
        model.matching.type: atss
        model.matching.topk: 9
        note: "ATSS with k=9 (SCRFD default)"
      - name: "atss_k5"
        model.matching.type: atss
        model.matching.topk: 5
        note: "ATSS with smaller k"
      - name: "iou_035"
        model.matching.type: iou
        model.matching.pos_iou: 0.35
        model.matching.neg_iou: 0.3
        note: "Fixed IoU (TinaFace-style lower threshold)"
      - name: "iou_050"
        model.matching.type: iou
        model.matching.pos_iou: 0.5
        model.matching.neg_iou: 0.4
        note: "Standard IoU threshold"

  # 4. Robustness Augmentations
  robustness:
    description: "Impact of blur/compression/lighting augmentations"
    experiments:
      - name: "no_robustness"
        data.robustness_augmentation: false
        note: "Standard aug only"
      - name: "blur_only"
        data.robustness_augmentation: true
        data.robustness_blur_prob: 0.3
        data.robustness_jpeg_prob: 0.0
        data.robustness_lowlight_prob: 0.0
        note: "Blur augmentation only"
      - name: "compression_only"
        data.robustness_augmentation: true
        data.robustness_blur_prob: 0.0
        data.robustness_jpeg_prob: 0.3
        data.robustness_lowlight_prob: 0.0
        note: "JPEG compression only"
      - name: "full_robustness"
        data.robustness_augmentation: true
        note: "All robustness augmentations"

  # 5. GroupNorm vs BatchNorm
  normalization:
    description: "GN vs BN (TinaFace found GN better at small batch sizes)"
    experiments:
      - name: "group_norm"
        model.neck.use_gn: true
        model.head.use_gn: true
        note: "GroupNorm (SCRFD/TinaFace default)"
      - name: "batch_norm"
        model.neck.use_gn: false
        model.head.use_gn: false
        note: "Standard BatchNorm"

  # 6. Input Resolution
  input_resolution:
    description: "Trade-off between resolution and speed for tiny faces"
    experiments:
      - name: "res_320"
        data.input_size: 320
        inference.input_size: 320
      - name: "res_480"
        data.input_size: 480
        inference.input_size: 480
      - name: "res_640"
        data.input_size: 640
        inference.input_size: 640
      - name: "res_960"
        data.input_size: 960
        inference.input_size: 960
        note: "Higher res for tiny faces (+5-10% Hard AP, ~4x slower)"

  # 7. Landmark Multi-Task Learning
  landmarks:
    description: "RetinaFace finding: auxiliary landmark loss improves detection"
    experiments:
      - name: "no_landmarks"
        model.head.use_landmarks: false
      - name: "with_landmarks"
        model.head.use_landmarks: true
        model.losses.lmk_weight: 0.1
        note: "Requires RetinaFace landmark annotations"

  # 8. Tracker Configuration
  tracking:
    description: "ByteTrack hyperparameter sensitivity"
    experiments:
      - name: "no_tracking"
        engine.use_tracking: false
      - name: "conservative"
        engine.tracker.high_thresh: 0.6
        engine.tracker.low_thresh: 0.2
        engine.tracker.max_lost: 15
      - name: "aggressive"
        engine.tracker.high_thresh: 0.3
        engine.tracker.low_thresh: 0.05
        engine.tracker.max_lost: 60
        note: "More tracks, more false positives"