sidhtang's picture
Update app.py
bd5362c verified
import cv2
import numpy as np
import mediapipe as mp
import torch
import torch.nn as nn
import torchvision.transforms as transforms
from PIL import Image
import gradio as gr
from enum import Enum
import colorsys
from typing import Tuple, Dict
import torch.nn.functional as F
class ClothingType(Enum):
SHIRT = "shirt"
PANTS = "pants"
DRESS = "dress"
JACKET = "jacket"
class BodySegmentation(nn.Module):
def __init__(self):
super().__init__()
# Load DeepLab v3+ for semantic segmentation
self.model = torch.hub.load('pytorch/vision:v0.10.0', 'deeplabv3_resnet50', pretrained=True)
self.model.eval()
def forward(self, x):
return self.model(x)['out']
class VirtualTryOn:
def __init__(self):
# Initialize MediaPipe
self.mp_pose = mp.solutions.pose
self.mp_holistic = mp.solutions.holistic
self.pose = self.mp_pose.Pose(
static_image_mode=True,
model_complexity=2,
min_detection_confidence=0.5
)
self.holistic = self.mp_holistic.Holistic(
static_image_mode=True,
model_complexity=2,
min_detection_confidence=0.5
)
# Initialize body segmentation
self.segmentation = BodySegmentation()
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.segmentation.to(self.device)
# Image transforms
self.transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
def get_body_segmentation(self, image: np.ndarray) -> np.ndarray:
"""
Get precise body segmentation mask
"""
# Prepare image for model
pil_image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
input_tensor = self.transforms(pil_image).unsqueeze(0).to(self.device)
# Get segmentation mask
with torch.no_grad():
output = self.segmentation(input_tensor)
mask = torch.argmax(output, dim=1).squeeze().cpu().numpy()
# Person class is typically index 15 in COCO dataset
return (mask == 15).astype(np.uint8)
def estimate_lighting(self, image: np.ndarray) -> Dict[str, float]:
"""
Estimate lighting conditions from the image
"""
# Convert to HSV
hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
# Get average brightness and saturation
brightness = np.mean(hsv[:, :, 2])
saturation = np.mean(hsv[:, :, 1])
return {
'brightness': brightness / 255.0,
'saturation': saturation / 255.0
}
def adjust_clothing_color(self, clothing: np.ndarray,
lighting_params: Dict[str, float]) -> np.ndarray:
"""
Adjust clothing colors to match lighting conditions
"""
# Convert to HSV for easier adjustment
hsv = cv2.cvtColor(clothing, cv2.COLOR_BGR2HSV).astype(np.float32)
# Adjust brightness and saturation
hsv[:, :, 2] *= lighting_params['brightness']
hsv[:, :, 1] *= lighting_params['saturation']
# Ensure values are within valid range
hsv = np.clip(hsv, 0, 255).astype(np.uint8)
# Convert back to BGR
return cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
def get_clothing_dimensions(self, landmarks, image_shape: Tuple[int, int],
clothing_type: ClothingType) -> Dict:
"""
Get clothing dimensions based on body landmarks and clothing type
"""
height, width = image_shape[:2]
if clothing_type in [ClothingType.SHIRT, ClothingType.JACKET]:
# For upper body clothing
left_shoulder = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_SHOULDER]
right_shoulder = landmarks.landmark[self.mp_pose.PoseLandmark.RIGHT_SHOULDER]
left_hip = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_HIP]
shoulder_width = abs(right_shoulder.x - left_shoulder.x) * width
torso_height = abs(left_shoulder.y - left_hip.y) * height
return {
'top_left': (
int(min(left_shoulder.x, right_shoulder.x) * width),
int(left_shoulder.y * height)
),
'width': int(shoulder_width * 1.3),
'height': int(torso_height * 1.1)
}
elif clothing_type == ClothingType.PANTS:
# For pants
left_hip = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_HIP]
right_hip = landmarks.landmark[self.mp_pose.PoseLandmark.RIGHT_HIP]
left_ankle = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_ANKLE]
hip_width = abs(right_hip.x - left_hip.x) * width
leg_height = abs(left_hip.y - left_ankle.y) * height
return {
'top_left': (
int(min(left_hip.x, right_hip.x) * width),
int(left_hip.y * height)
),
'width': int(hip_width * 1.5),
'height': int(leg_height * 1.05)
}
elif clothing_type == ClothingType.DRESS:
# For dresses
left_shoulder = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_SHOULDER]
right_shoulder = landmarks.landmark[self.mp_pose.PoseLandmark.RIGHT_SHOULDER]
left_knee = landmarks.landmark[self.mp_pose.PoseLandmark.LEFT_KNEE]
shoulder_width = abs(right_shoulder.x - left_shoulder.x) * width
dress_height = abs(left_shoulder.y - left_knee.y) * height
return {
'top_left': (
int(min(left_shoulder.x, right_shoulder.x) * width),
int(left_shoulder.y * height)
),
'width': int(shoulder_width * 1.4),
'height': int(dress_height * 1.1)
}
def try_on(self, person_image: np.ndarray, clothing_image: np.ndarray,
clothing_type: ClothingType) -> np.ndarray:
"""
Enhanced try-on method with support for different clothing types
"""
# Get body segmentation
body_mask = self.get_body_segmentation(person_image)
# Get pose landmarks
results = self.pose.process(cv2.cvtColor(person_image, cv2.COLOR_BGR2RGB))
if not results.pose_landmarks:
raise ValueError("No person detected in the image")
# Estimate lighting conditions
lighting_params = self.estimate_lighting(person_image)
# Adjust clothing colors
adjusted_clothing = self.adjust_clothing_color(clothing_image, lighting_params)
# Get clothing dimensions
dimensions = self.get_clothing_dimensions(
results.pose_landmarks,
person_image.shape,
clothing_type
)
# Resize clothing
clothing_resized = cv2.resize(
adjusted_clothing,
(dimensions['width'], dimensions['height']),
interpolation=cv2.INTER_AREA
)
# Create alpha mask for smooth blending
if clothing_resized.shape[2] == 4:
alpha_channel = clothing_resized[:, :, 3] / 255.0
else:
alpha_channel = np.ones(clothing_resized.shape[:2])
alpha_3channel = np.stack([alpha_channel] * 3, axis=2)
# Calculate placement coordinates
y1 = dimensions['top_left'][1]
y2 = y1 + dimensions['height']
x1 = dimensions['top_left'][0]
x2 = x1 + dimensions['width']
# Ensure coordinates are within image boundaries
y1 = max(0, y1)
y2 = min(person_image.shape[0], y2)
x1 = max(0, x1)
x2 = min(person_image.shape[1], x2)
# Apply body mask to improve blending
body_mask_roi = body_mask[y1:y2, x1:x2]
alpha_3channel = alpha_3channel * np.expand_dims(body_mask_roi, axis=2)
# Blend images
roi = person_image[y1:y2, x1:x2]
clothing_rgb = clothing_resized[:, :, :3]
blended = (1 - alpha_3channel) * roi + alpha_3channel * clothing_rgb[:roi.shape[0], :roi.shape[1]]
result = person_image.copy()
result[y1:y2, x1:x2] = blended
return result
def create_gradio_interface():
def process_images(person_img, clothing_img, clothing_type):
try_on = VirtualTryOn()
# Convert clothing type string to enum
clothing_type_enum = ClothingType(clothing_type.lower())
# Process the images
result = try_on.try_on(person_img, clothing_img, clothing_type_enum)
return result
# Create the interface
iface = gr.Interface(
fn=process_images,
inputs=[
gr.Image(label="Upload Person Image"),
gr.Image(label="Upload Clothing Image"),
gr.Dropdown(
choices=["Shirt", "Pants", "Dress", "Jacket"],
label="Select Clothing Type"
)
],
outputs=gr.Image(label="Result"),
title="Virtual Try-On System",
description="Upload a person's image and a clothing item to see how it looks!",
examples=[
["person.jpg", "shirt.png", "Shirt"],
["person.jpg", "pants.png", "Pants"]
]
)
return iface
if __name__ == "__main__":
iface = create_gradio_interface()
iface.launch()