File size: 6,663 Bytes
afaa2cf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 | """
SCRFD Detection Head — shared-weight, multi-task, scale-aware.
Design from SCRFD paper:
- Weight sharing across pyramid levels (parameter-efficient)
- GroupNorm for batch-size independence
- Separate cls and reg branches (GFL-style)
- Optional landmark branch (RetinaFace-style 5-point)
Output per anchor:
- Classification: 1 score (face quality score via GFL)
- Box regression: 4 values (distance from anchor center to box edges)
- Landmarks (optional): 10 values (5 x,y offsets from anchor center)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import List, Tuple, Optional
import math
class SCRFDHead(nn.Module):
"""
Shared detection head applied to each FPN level.
Args:
in_channels: Input channels from neck
num_classes: Number of classes (1 for face detection)
num_anchors: Anchors per spatial location per level
feat_channels: Hidden channel width in head convolutions
stacked_convs: Number of stacked 3×3 convs in each branch
use_gn: Use GroupNorm (vs BatchNorm)
use_landmarks: Enable 5-point landmark regression branch
"""
def __init__(self,
in_channels: int = 64,
num_classes: int = 1,
num_anchors: int = 2,
feat_channels: int = 64,
stacked_convs: int = 2,
use_gn: bool = True,
use_landmarks: bool = False):
super().__init__()
self.num_classes = num_classes
self.num_anchors = num_anchors
self.use_landmarks = use_landmarks
# Classification branch
cls_convs = []
for i in range(stacked_convs):
ch_in = in_channels if i == 0 else feat_channels
cls_convs.append(nn.Conv2d(ch_in, feat_channels, 3, 1, 1, bias=False))
if use_gn:
gn_groups = min(16, feat_channels)
while feat_channels % gn_groups != 0:
gn_groups -= 1
cls_convs.append(nn.GroupNorm(gn_groups, feat_channels))
else:
cls_convs.append(nn.BatchNorm2d(feat_channels))
cls_convs.append(nn.ReLU(inplace=True))
self.cls_convs = nn.Sequential(*cls_convs)
self.cls_out = nn.Conv2d(feat_channels, num_anchors * num_classes, 3, 1, 1)
# Box regression branch
reg_convs = []
for i in range(stacked_convs):
ch_in = in_channels if i == 0 else feat_channels
reg_convs.append(nn.Conv2d(ch_in, feat_channels, 3, 1, 1, bias=False))
if use_gn:
gn_groups = min(16, feat_channels)
while feat_channels % gn_groups != 0:
gn_groups -= 1
reg_convs.append(nn.GroupNorm(gn_groups, feat_channels))
else:
reg_convs.append(nn.BatchNorm2d(feat_channels))
reg_convs.append(nn.ReLU(inplace=True))
self.reg_convs = nn.Sequential(*reg_convs)
self.reg_out = nn.Conv2d(feat_channels, num_anchors * 4, 3, 1, 1)
# Landmark branch (optional)
if use_landmarks:
lmk_convs = []
for i in range(stacked_convs):
ch_in = in_channels if i == 0 else feat_channels
lmk_convs.append(nn.Conv2d(ch_in, feat_channels, 3, 1, 1, bias=False))
if use_gn:
gn_groups = min(16, feat_channels)
while feat_channels % gn_groups != 0:
gn_groups -= 1
lmk_convs.append(nn.GroupNorm(gn_groups, feat_channels))
else:
lmk_convs.append(nn.BatchNorm2d(feat_channels))
lmk_convs.append(nn.ReLU(inplace=True))
self.lmk_convs = nn.Sequential(*lmk_convs)
self.lmk_out = nn.Conv2d(feat_channels, num_anchors * 10, 3, 1, 1)
self._init_weights()
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.normal_(m.weight, std=0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Initialize cls bias for focal loss (prevents initial instability)
# Prior probability = 0.01
prior_prob = 0.01
bias_init = -math.log((1 - prior_prob) / prior_prob)
nn.init.constant_(self.cls_out.bias, bias_init)
def forward_single(self, x: torch.Tensor) -> dict:
"""Forward pass for a single FPN level."""
cls_feat = self.cls_convs(x)
cls_score = self.cls_out(cls_feat) # [B, A*C, H, W]
reg_feat = self.reg_convs(x)
bbox_pred = self.reg_out(reg_feat) # [B, A*4, H, W]
result = {'cls_score': cls_score, 'bbox_pred': bbox_pred}
if self.use_landmarks:
lmk_feat = self.lmk_convs(x)
lmk_pred = self.lmk_out(lmk_feat) # [B, A*10, H, W]
result['lmk_pred'] = lmk_pred
return result
def forward(self, features: Tuple[torch.Tensor, ...]) -> dict:
"""
Forward on all FPN levels.
Args:
features: (P3, P4, P5) from neck
Returns:
dict with keys 'cls_scores', 'bbox_preds', optionally 'lmk_preds'
Each value is a list of tensors, one per level.
"""
cls_scores = []
bbox_preds = []
lmk_preds = []
for feat in features:
out = self.forward_single(feat)
cls_scores.append(out['cls_score'])
bbox_preds.append(out['bbox_pred'])
if self.use_landmarks:
lmk_preds.append(out['lmk_pred'])
result = {'cls_scores': cls_scores, 'bbox_preds': bbox_preds}
if self.use_landmarks:
result['lmk_preds'] = lmk_preds
return result
# ──────────────────────── Configuration presets ────────────────────────
HEAD_CONFIGS = {
'scrfd_34g': dict(feat_channels=64, stacked_convs=3),
'scrfd_10g': dict(feat_channels=56, stacked_convs=2),
'scrfd_2.5g': dict(feat_channels=40, stacked_convs=2),
'scrfd_0.5g': dict(feat_channels=16, stacked_convs=2),
}
def build_head(name: str, in_channels: int, **kwargs) -> SCRFDHead:
"""Build detection head by model name."""
cfg = HEAD_CONFIGS.get(name, {})
cfg.update(kwargs)
return SCRFDHead(in_channels=in_channels, **cfg)
|