face-detection / scrfd.py
orik-ss's picture
Face detection Space: dual SCRFD (500m + 2.5g) side-by-side comparison
cb68e53
Raw
History Blame Contribute Delete
8.18 kB
"""Standalone SCRFD face detector (ONNX Runtime, CPU).
A self-contained re-implementation of InsightFace's SCRFD post-processing so the
app can load a raw ``.onnx`` file directly — no ``insightface`` package, no model
pack directory. Handles the 6/9/10/15-output SCRFD variants; ``det_500m.onnx`` and
``det_2.5g.onnx`` are both 9-output (3 strides x {score, bbox, kps}) with 5-point
landmarks.
Reference: https://github.com/deepinsight/insightface (scrfd.py).
"""
import cv2
import numpy as np
import onnxruntime
def distance2bbox(points, distance):
"""Decode (left, top, right, bottom) distances from an anchor center to a box."""
x1 = points[:, 0] - distance[:, 0]
y1 = points[:, 1] - distance[:, 1]
x2 = points[:, 0] + distance[:, 2]
y2 = points[:, 1] + distance[:, 3]
return np.stack([x1, y1, x2, y2], axis=-1)
def distance2kps(points, distance):
"""Decode per-keypoint (dx, dy) distances from an anchor center to landmarks."""
preds = []
for i in range(0, distance.shape[1], 2):
px = points[:, i % 2] + distance[:, i]
py = points[:, i % 2 + 1] + distance[:, i + 1]
preds.append(px)
preds.append(py)
return np.stack(preds, axis=-1)
class SCRFD:
"""SCRFD ONNX face detector.
Parameters
----------
model_file : str
Path to the ``.onnx`` file.
providers : list[str] | None
ONNX Runtime execution providers. Defaults to CPU.
"""
def __init__(self, model_file, providers=None):
self.model_file = model_file
providers = providers or ["CPUExecutionProvider"]
self.session = onnxruntime.InferenceSession(model_file, providers=providers)
self.center_cache = {}
self.nms_thresh = 0.4
self.input_mean = 127.5
self.input_std = 128.0
self._init_vars()
def _init_vars(self):
inp = self.session.get_inputs()[0]
self.input_name = inp.name
# Static input size if the model declares one (e.g. [1, 3, 640, 640]),
# else fall back to 640x640 at detect() time.
shape = inp.shape
if isinstance(shape[2], int) and isinstance(shape[3], int):
self.input_size = (shape[3], shape[2]) # (w, h)
else:
self.input_size = (640, 640)
outputs = self.session.get_outputs()
self.output_names = [o.name for o in outputs]
self.use_kps = False
self._num_anchors = 1
n = len(outputs)
if n == 6:
self.fmc, self._feat_stride_fpn, self._num_anchors = 3, [8, 16, 32], 2
elif n == 9:
self.fmc, self._feat_stride_fpn, self._num_anchors = 3, [8, 16, 32], 2
self.use_kps = True
elif n == 10:
self.fmc, self._feat_stride_fpn, self._num_anchors = 5, [8, 16, 32, 64, 128], 1
elif n == 15:
self.fmc, self._feat_stride_fpn, self._num_anchors = 5, [8, 16, 32, 64, 128], 1
self.use_kps = True
else:
raise ValueError(f"Unexpected SCRFD output count: {n}")
def forward(self, img, thresh):
scores_list, bboxes_list, kpss_list = [], [], []
blob = cv2.dnn.blobFromImage(
img,
1.0 / self.input_std,
(img.shape[1], img.shape[0]),
(self.input_mean, self.input_mean, self.input_mean),
swapRB=True,
)
net_outs = self.session.run(self.output_names, {self.input_name: blob})
input_height, input_width = blob.shape[2], blob.shape[3]
fmc = self.fmc
for idx, stride in enumerate(self._feat_stride_fpn):
scores = net_outs[idx]
bbox_preds = net_outs[idx + fmc] * stride
if self.use_kps:
kps_preds = net_outs[idx + fmc * 2] * stride
height, width = input_height // stride, input_width // stride
key = (height, width, stride)
if key in self.center_cache:
anchor_centers = self.center_cache[key]
else:
anchor_centers = np.stack(
np.mgrid[:height, :width][::-1], axis=-1
).astype(np.float32)
anchor_centers = (anchor_centers * stride).reshape((-1, 2))
if self._num_anchors > 1:
anchor_centers = np.stack(
[anchor_centers] * self._num_anchors, axis=1
).reshape((-1, 2))
if len(self.center_cache) < 100:
self.center_cache[key] = anchor_centers
pos_inds = np.where(scores >= thresh)[0]
bboxes = distance2bbox(anchor_centers, bbox_preds)
scores_list.append(scores[pos_inds])
bboxes_list.append(bboxes[pos_inds])
if self.use_kps:
kpss = distance2kps(anchor_centers, kps_preds)
kpss = kpss.reshape((kpss.shape[0], -1, 2))
kpss_list.append(kpss[pos_inds])
return scores_list, bboxes_list, kpss_list
def nms(self, dets):
"""Standard IoU NMS on ``[x1, y1, x2, y2, score]`` rows (score-sorted)."""
x1, y1, x2, y2, scores = dets[:, 0], dets[:, 1], dets[:, 2], dets[:, 3], dets[:, 4]
areas = (x2 - x1 + 1) * (y2 - y1 + 1)
order = scores.argsort()[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
xx1 = np.maximum(x1[i], x1[order[1:]])
yy1 = np.maximum(y1[i], y1[order[1:]])
xx2 = np.minimum(x2[i], x2[order[1:]])
yy2 = np.minimum(y2[i], y2[order[1:]])
w = np.maximum(0.0, xx2 - xx1 + 1)
h = np.maximum(0.0, yy2 - yy1 + 1)
inter = w * h
ovr = inter / (areas[i] + areas[order[1:]] - inter)
inds = np.where(ovr <= self.nms_thresh)[0]
order = order[inds + 1]
return keep
def detect(self, img, thresh=0.5, input_size=None, max_num=0, metric="default"):
"""Detect faces in a BGR ``uint8`` image.
Returns ``(dets, kpss)`` where ``dets`` is ``[N, 5]`` (x1,y1,x2,y2,score)
in original-image pixels and ``kpss`` is ``[N, 5, 2]`` landmarks (or None).
"""
input_size = self.input_size if input_size is None else input_size
# Letterbox: preserve aspect ratio, pad to the model's square input.
im_ratio = float(img.shape[0]) / img.shape[1]
model_ratio = float(input_size[1]) / input_size[0]
if im_ratio > model_ratio:
new_height = input_size[1]
new_width = int(new_height / im_ratio)
else:
new_width = input_size[0]
new_height = int(new_width * im_ratio)
det_scale = float(new_height) / img.shape[0]
resized = cv2.resize(img, (new_width, new_height))
det_img = np.zeros((input_size[1], input_size[0], 3), dtype=np.uint8)
det_img[:new_height, :new_width, :] = resized
scores_list, bboxes_list, kpss_list = self.forward(det_img, thresh)
scores = np.vstack(scores_list)
order = scores.ravel().argsort()[::-1]
bboxes = np.vstack(bboxes_list) / det_scale
pre_det = np.hstack((bboxes, scores)).astype(np.float32, copy=False)
pre_det = pre_det[order, :]
keep = self.nms(pre_det)
det = pre_det[keep, :]
kpss = None
if self.use_kps:
kpss = np.vstack(kpss_list) / det_scale
kpss = kpss[order, :, :][keep, :, :]
if max_num > 0 and det.shape[0] > max_num:
area = (det[:, 2] - det[:, 0]) * (det[:, 3] - det[:, 1])
img_center = img.shape[0] // 2, img.shape[1] // 2
offsets = np.vstack([
(det[:, 0] + det[:, 2]) / 2 - img_center[1],
(det[:, 1] + det[:, 3]) / 2 - img_center[0],
])
offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
values = area if metric == "max" else area - offset_dist_squared * 2.0
bindex = np.argsort(values)[::-1][:max_num]
det = det[bindex, :]
if kpss is not None:
kpss = kpss[bindex, :, :]
return det, kpss