rafiaashraf's picture
Update app.py
b4e6d93 verified
raw
history blame
16.9 kB
import cv2
import numpy as np
import mediapipe as mp
import torch
from flask import Flask, request, jsonify
import torch.nn.functional as F
from flask_cors import CORS
import os
import shutil
# --- Monkey Patch: Override model path before using mediapipe.pose ---
from mediapipe.python.solutions import pose as mp_pose_module
CUSTOM_MODEL_PATH = "/app/models/pose_landmark_heavy.tflite"
def monkey_patch_pose_path(model_complexity):
model_complexity_map = {
0: "pose_landmark_lite",
1: "pose_landmark_full",
2: "pose_landmark_heavy"
}
if model_complexity in model_complexity_map:
file_name = model_complexity_map[model_complexity] + ".tflite"
mp_pose_module.MODEL_PATHS[model_complexity] = CUSTOM_MODEL_PATH
else:
raise ValueError(f"Unsupported model complexity: {model_complexity}")
# Monkey patch before creating Pose instance
monkey_patch_pose_path(model_complexity=2)
app = Flask(__name__)
CORS(app)
mp_pose = mp.solutions.pose
mp_holistic = mp.solutions.holistic
pose = mp_pose.Pose(
model_complexity=2,
static_image_mode=True,
min_detection_confidence=0.5,
min_tracking_confidence=0.5
) # Improved accuracy
holistic = mp_holistic.Holistic() # For refining pose
KNOWN_OBJECT_WIDTH_CM = 21.0 # A4 paper width in cm
FOCAL_LENGTH = 600 # Default focal length
DEFAULT_HEIGHT_CM = 152.0 # Default height if not provided
# Load depth estimation model
def load_depth_model():
model = torch.hub.load("intel-isl/MiDaS", "MiDaS_small")
model.eval()
return model
depth_model = load_depth_model()
def calibrate_focal_length(image, real_width_cm, detected_width_px):
"""Dynamically calibrates focal length using a known object."""
return (detected_width_px * FOCAL_LENGTH) / real_width_cm if detected_width_px else FOCAL_LENGTH
def detect_reference_object(image):
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
edges = cv2.Canny(gray, 50, 150)
contours, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if contours:
largest_contour = max(contours, key=cv2.contourArea)
x, y, w, h = cv2.boundingRect(largest_contour)
focal_length = calibrate_focal_length(image, KNOWN_OBJECT_WIDTH_CM, w)
scale_factor = KNOWN_OBJECT_WIDTH_CM / w
return scale_factor, focal_length
return 0.05, FOCAL_LENGTH
def estimate_depth(image):
"""Uses AI-based depth estimation to improve circumference calculations."""
input_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) / 255.0
input_tensor = torch.tensor(input_image, dtype=torch.float32).permute(2, 0, 1).unsqueeze(0)
# Resize input to match MiDaS model input size
input_tensor = F.interpolate(input_tensor, size=(384, 384), mode="bilinear", align_corners=False)
with torch.no_grad():
depth_map = depth_model(input_tensor)
return depth_map.squeeze().numpy()
def calculate_distance_using_height(landmarks, image_height, user_height_cm):
"""Calculate distance using the user's known height."""
top_head = landmarks[mp_pose.PoseLandmark.NOSE.value].y * image_height
bottom_foot = max(
landmarks[mp_pose.PoseLandmark.LEFT_ANKLE.value].y,
landmarks[mp_pose.PoseLandmark.RIGHT_ANKLE.value].y
) * image_height
person_height_px = abs(bottom_foot - top_head)
# Using the formula: distance = (actual_height_cm * focal_length) / height_in_pixels
distance = (user_height_cm * FOCAL_LENGTH) / person_height_px
# Calculate more accurate scale_factor based on known height
scale_factor = user_height_cm / person_height_px
return distance, scale_factor
def get_body_width_at_height(frame, height_px, center_x):
"""Scan horizontally at a specific height to find body edges."""
# Convert to grayscale and apply threshold
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
_, thresh = cv2.threshold(blur, 50, 255, cv2.THRESH_BINARY)
# Ensure height_px is within image bounds
if height_px >= frame.shape[0]:
height_px = frame.shape[0] - 1
# Get horizontal line at the specified height
horizontal_line = thresh[height_px, :]
# Find left and right edges starting from center
center_x = int(center_x * frame.shape[1])
left_edge, right_edge = center_x, center_x
# Scan from center to left
for i in range(center_x, 0, -1):
if horizontal_line[i] == 0: # Found edge (black pixel)
left_edge = i
break
# Scan from center to right
for i in range(center_x, len(horizontal_line)):
if horizontal_line[i] == 0: # Found edge (black pixel)
right_edge = i
break
width_px = right_edge - left_edge
# If width is unreasonably small, apply a minimum width
min_width = 0.1 * frame.shape[1] # Minimum width as 10% of image width
if width_px < min_width:
width_px = min_width
return width_px
def calculate_measurements(results, scale_factor, image_width, image_height, depth_map, frame=None, user_height_cm=None):
landmarks = results.pose_landmarks.landmark
# If user's height is provided, use it to get a more accurate scale factor
if user_height_cm:
_, scale_factor = calculate_distance_using_height(landmarks, image_height, user_height_cm)
def pixel_to_cm(value):
return round(value * scale_factor, 2)
def calculate_circumference(width_px, depth_ratio=1.0):
"""Estimate circumference using width and depth adjustment."""
# Using a simplified elliptical approximation: C ≈ 2π * sqrt((a² + b²)/2)
# where a is half the width and b is estimated depth
width_cm = width_px * scale_factor
estimated_depth_cm = width_cm * depth_ratio * 0.7 # Depth is typically ~70% of width for torso
half_width = width_cm / 2
half_depth = estimated_depth_cm / 2
return round(2 * np.pi * np.sqrt((half_width**2 + half_depth**2) / 2), 2)
measurements = {}
# Shoulder Width
left_shoulder = landmarks[mp_pose.PoseLandmark.LEFT_SHOULDER.value]
right_shoulder = landmarks[mp_pose.PoseLandmark.RIGHT_SHOULDER.value]
shoulder_width_px = abs(left_shoulder.x * image_width - right_shoulder.x * image_width)
# Apply a slight correction factor for shoulders (they're usually detected well)
shoulder_correction = 1.1 # 10% wider
shoulder_width_px *= shoulder_correction
measurements["shoulder_width"] = pixel_to_cm(shoulder_width_px)
# Chest/Bust Measurement
chest_y_ratio = 0.15 # Approximately 15% down from shoulder to hip
chest_y = left_shoulder.y + (landmarks[mp_pose.PoseLandmark.LEFT_HIP.value].y - left_shoulder.y) * chest_y_ratio
chest_correction = 1.15 # 15% wider than detected width
chest_width_px = abs((right_shoulder.x - left_shoulder.x) * image_width) * chest_correction
if frame is not None:
chest_y_px = int(chest_y * image_height)
center_x = (left_shoulder.x + right_shoulder.x) / 2
detected_width = get_body_width_at_height(frame, chest_y_px, center_x)
if detected_width > 0:
chest_width_px = max(chest_width_px, detected_width)
chest_depth_ratio = 1.0
if depth_map is not None:
chest_x = int(((left_shoulder.x + right_shoulder.x) / 2) * image_width)
chest_y_px = int(chest_y * image_height)
scale_y = 384 / image_height
scale_x = 384 / image_width
chest_y_scaled = int(chest_y_px * scale_y)
chest_x_scaled = int(chest_x * scale_x)
if 0 <= chest_y_scaled < 384 and 0 <= chest_x_scaled < 384:
chest_depth = depth_map[chest_y_scaled, chest_x_scaled]
max_depth = np.max(depth_map)
chest_depth_ratio = 1.0 + 0.5 * (1.0 - chest_depth / max_depth)
measurements["chest_width"] = pixel_to_cm(chest_width_px)
measurements["chest_circumference"] = calculate_circumference(chest_width_px, chest_depth_ratio)
# Waist Measurement
left_hip = landmarks[mp_pose.PoseLandmark.LEFT_HIP.value]
right_hip = landmarks[mp_pose.PoseLandmark.RIGHT_HIP.value]
# Adjust waist_y_ratio to better reflect the natural waistline
waist_y_ratio = 0.35 # 35% down from shoulder to hip (higher than before)
waist_y = left_shoulder.y + (left_hip.y - left_shoulder.y) * waist_y_ratio
# Use contour detection to dynamically estimate waist width
if frame is not None:
waist_y_px = int(waist_y * image_height)
center_x = (left_hip.x + right_hip.x) / 2
detected_width = get_body_width_at_height(frame, waist_y_px, center_x)
if detected_width > 0:
waist_width_px = detected_width
else:
# Fallback to hip width if contour detection fails
waist_width_px = abs(right_hip.x - left_hip.x) * image_width * 0.9 # 90% of hip width
else:
# Fallback to hip width if no frame is provided
waist_width_px = abs(right_hip.x - left_hip.x) * image_width * 0.9 # 90% of hip width
# Apply 30% correction factor to waist width
waist_correction = 1.16 # 30% wider
waist_width_px *= waist_correction
# Get depth adjustment for waist if available
waist_depth_ratio = 1.0
if depth_map is not None:
waist_x = int(((left_hip.x + right_hip.x) / 2) * image_width)
waist_y_px = int(waist_y * image_height)
scale_y = 384 / image_height
scale_x = 384 / image_width
waist_y_scaled = int(waist_y_px * scale_y)
waist_x_scaled = int(waist_x * scale_x)
if 0 <= waist_y_scaled < 384 and 0 <= waist_x_scaled < 384:
waist_depth = depth_map[waist_y_scaled, waist_x_scaled]
max_depth = np.max(depth_map)
waist_depth_ratio = 1.0 + 0.5 * (1.0 - waist_depth / max_depth)
measurements["waist_width"] = pixel_to_cm(waist_width_px)
measurements["waist"] = calculate_circumference(waist_width_px, waist_depth_ratio)
# Hip Measurement
hip_correction = 1.35 # Hips are typically 35% wider than detected landmarks
hip_width_px = abs(left_hip.x * image_width - right_hip.x * image_width) * hip_correction
if frame is not None:
hip_y_offset = 0.1 # 10% down from hip landmarks
hip_y = left_hip.y + (landmarks[mp_pose.PoseLandmark.LEFT_KNEE.value].y - left_hip.y) * hip_y_offset
hip_y_px = int(hip_y * image_height)
center_x = (left_hip.x + right_hip.x) / 2
detected_width = get_body_width_at_height(frame, hip_y_px, center_x)
if detected_width > 0:
hip_width_px = max(hip_width_px, detected_width)
hip_depth_ratio = 1.0
if depth_map is not None:
hip_x = int(((left_hip.x + right_hip.x) / 2) * image_width)
hip_y_px = int(left_hip.y * image_height)
hip_y_scaled = int(hip_y_px * scale_y)
hip_x_scaled = int(hip_x * scale_x)
if 0 <= hip_y_scaled < 384 and 0 <= hip_x_scaled < 384:
hip_depth = depth_map[hip_y_scaled, hip_x_scaled]
max_depth = np.max(depth_map)
hip_depth_ratio = 1.0 + 0.5 * (1.0 - hip_depth / max_depth)
measurements["hip_width"] = pixel_to_cm(hip_width_px)
measurements["hip"] = calculate_circumference(hip_width_px, hip_depth_ratio)
# Other measurements (unchanged)
neck = landmarks[mp_pose.PoseLandmark.NOSE.value]
left_ear = landmarks[mp_pose.PoseLandmark.LEFT_EAR.value]
neck_width_px = abs(neck.x * image_width - left_ear.x * image_width) * 2.0
measurements["neck"] = calculate_circumference(neck_width_px, 1.0)
measurements["neck_width"] = pixel_to_cm(neck_width_px)
left_wrist = landmarks[mp_pose.PoseLandmark.LEFT_WRIST.value]
sleeve_length_px = abs(left_shoulder.y * image_height - left_wrist.y * image_height)
measurements["arm_length"] = pixel_to_cm(sleeve_length_px)
shirt_length_px = abs(left_shoulder.y * image_height - left_hip.y * image_height) * 1.2
measurements["shirt_length"] = pixel_to_cm(shirt_length_px)
# Thigh Circumference (improved with depth information)
thigh_y_ratio = 0.2 # 20% down from hip to knee
left_knee = landmarks[mp_pose.PoseLandmark.LEFT_KNEE.value]
thigh_y = left_hip.y + (left_knee.y - left_hip.y) * thigh_y_ratio
# Apply correction factor for thigh width
thigh_correction = 1.2 # Thighs are typically wider than what can be estimated from front view
thigh_width_px = hip_width_px * 0.5 * thigh_correction # Base thigh width on hip width
# Use contour detection if frame is available
if frame is not None:
thigh_y_px = int(thigh_y * image_height)
thigh_x = left_hip.x * 0.9 # Move slightly inward from hip
detected_width = get_body_width_at_height(frame, thigh_y_px, thigh_x)
if detected_width > 0 and detected_width < hip_width_px: # Sanity check
thigh_width_px = detected_width # Use detected width
# If depth map is available, use it for thigh measurement
thigh_depth_ratio = 1.0
if depth_map is not None:
thigh_x = int(left_hip.x * image_width)
thigh_y_px = int(thigh_y * image_height)
# Scale coordinates to match depth map size
thigh_y_scaled = int(thigh_y_px * scale_y)
thigh_x_scaled = int(thigh_x * scale_x)
if 0 <= thigh_y_scaled < 384 and 0 <= thigh_x_scaled < 384:
thigh_depth = depth_map[thigh_y_scaled, thigh_x_scaled]
max_depth = np.max(depth_map)
thigh_depth_ratio = 1.0 + 0.5 * (1.0 - thigh_depth / max_depth)
measurements["thigh"] = pixel_to_cm(thigh_width_px)
measurements["thigh_circumference"] = calculate_circumference(thigh_width_px, thigh_depth_ratio)
left_ankle = landmarks[mp_pose.PoseLandmark.LEFT_ANKLE.value]
trouser_length_px = abs(left_hip.y * image_height - left_ankle.y * image_height)
measurements["trouser_length"] = pixel_to_cm(trouser_length_px)
return measurements
@app.route("/upload_images", methods=["POST"])
def upload_images():
if "front" not in request.files:
return jsonify({"error": "Missing front image for reference."}), 400
# Get user height if provided, otherwise use default
user_height_cm = request.form.get('height_cm')
print(user_height_cm)
if user_height_cm:
try:
user_height_cm = float(user_height_cm)
except ValueError:
user_height_cm = DEFAULT_HEIGHT_CM
else:
user_height_cm = DEFAULT_HEIGHT_CM
received_images = {pose_name: request.files[pose_name] for pose_name in ["front", "left_side", "right_side", "back"] if pose_name in request.files}
measurements, scale_factor, focal_length, results = {}, None, FOCAL_LENGTH, {}
frames = {}
for pose_name, image_file in received_images.items():
image_np = np.frombuffer(image_file.read(), np.uint8)
frame = cv2.imdecode(image_np, cv2.IMREAD_COLOR)
frames[pose_name] = frame # Store the frame for contour detection
rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
results[pose_name] = holistic.process(rgb_frame)
image_height, image_width, _ = frame.shape
if pose_name == "front":
# Always use height for calibration (default or provided)
if results[pose_name].pose_landmarks:
_, scale_factor = calculate_distance_using_height(
results[pose_name].pose_landmarks.landmark,
image_height,
user_height_cm
)
else:
# Fallback to object detection only if pose landmarks aren't detected
scale_factor, focal_length = detect_reference_object(frame)
depth_map = estimate_depth(frame) if pose_name in ["front", "left_side"] else None
if results[pose_name].pose_landmarks:
if pose_name == "front":
measurements.update(calculate_measurements(
results[pose_name],
scale_factor,
image_width,
image_height,
depth_map,
frames[pose_name], # Pass the frame for contour detection
user_height_cm
))
# Debug information to help troubleshoot measurements
debug_info = {
"scale_factor": float(scale_factor) if scale_factor else None,
"focal_length": float(focal_length),
"user_height_cm": float(user_height_cm)
}
return jsonify({
"measurements": measurements,
"debug_info": debug_info
})
if __name__ == '__main__':
app.run(host='0.0.0.0', port=8000)