app.py

from io import BytesIO

import cv2
import gradio as gr
import numpy as np
import requests
from PIL import Image


from super_gradients.common.object_names import Models
from super_gradients.training import models
from super_gradients.training.utils.visualization.detection import draw_bbox
from super_gradients.training.utils.visualization.pose_estimation import PoseVisualization

# Initialize your pose estimation model
yolo_nas_pose = models.get("yolo_nas_pose_l",
                           num_classes=17,
                           checkpoint_path="/content/yolo_nas_pose_l_coco_pose.pth")

def process_and_predict(url=None,
                        image=None,
                        confidence=0.5,
                        iou=0.5):
    # If a URL is provided, use it directly for prediction
    if url is not None and url.strip() != "":
        response = requests.get(url)
        image = Image.open(BytesIO(response.content))
        image = np.array(image)
        result = yolo_nas_pose.predict(image, conf=confidence,iou=iou)
    # If a file is uploaded, read it, convert it to a numpy array and use it for prediction
    elif image is not None:
        result = yolo_nas_pose.predict(image, conf=confidence,iou=iou)
    else:
        return None  # If no input is provided, return None

    # Extract prediction data
    image_prediction = result._images_prediction_lst[0]

    pose_data = image_prediction.prediction

    # Visualize the prediction
    output_image = PoseVisualization.draw_poses(
        image=image_prediction.image,
        poses=pose_data.poses,
        boxes=pose_data.bboxes_xyxy,
        scores=pose_data.scores,
        is_crowd=None,
        edge_links=pose_data.edge_links,
        edge_colors=pose_data.edge_colors,
        keypoint_colors=pose_data.keypoint_colors,
        joint_thickness=2,
        box_thickness=2,
        keypoint_radius=5
    )

    blank_image = np.zeros_like(image_prediction.image)

    skeleton_image = PoseVisualization.draw_poses(
    image=blank_image,
    poses=pose_data.poses,
    boxes=pose_data.bboxes_xyxy,
    scores=pose_data.scores,
    is_crowd=None,
    edge_links=pose_data.edge_links,
    edge_colors=pose_data.edge_colors,
    keypoint_colors=pose_data.keypoint_colors,
    joint_thickness=2,
    box_thickness=2,
    keypoint_radius=5
)

    # Convert the resulting visualization to a PIL Image
    # output_image_pil = Image.fromarray(output_image.astype('uint8'), 'RGB')

    # Return the PIL Image directly
    return output_image, skeleton_image

# Define the Gradio interface
iface = gr.Interface(
    fn=process_and_predict,
    inputs=[
        gr.Textbox(placeholder="Enter Image URL", label="Image URL"),
        gr.Image(label="Upload Image", type='numpy'),
        gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label="Confidence Threshold"),
        gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label="IoU Threshold")
    ],
    outputs=[
        gr.components.Image(label="Estimated Pose"),
        gr.components.Image(label="Skeleton Only")
    ],
    title="YOLO-NAS-Pose Demo",
    description="Upload an image, enter an image URL, or use your webcam to use a pretrained YOLO-NAS-Pose L for inference. You can check out the ",
    live=False,
    allow_flagging=False,
    
    )

# Launch the interface
iface.launch()