app.py

from ultralytics import YOLO
import cv2
import gradio as gr
import numpy as np
import os
import torch
import utils
import plotly.graph_objects as go
import spaces

from image_segmenter import ImageSegmenter
from monocular_depth_estimator import MonocularDepthEstimator
from point_cloud_generator import display_pcd

# params
CANCEL_PROCESSING = False

# Initialize models (but actual loading happens in decorated functions)
img_seg = ImageSegmenter(model_type="yolov8s-seg")
depth_estimator = MonocularDepthEstimator(model_type="midas_v21_small_256")

@spaces.GPU(duration=30)  # Adjust duration based on your needs
def process_image(image):
    image = utils.resize(image)
    image_segmentation, objects_data = img_seg.predict(image)
    depthmap, depth_colormap = depth_estimator.make_prediction(image)
    dist_image = utils.draw_depth_info(image, depthmap, objects_data)
    objs_pcd = utils.generate_obj_pcd(depthmap, objects_data)
    plot_fig = display_pcd(objs_pcd)
    return image_segmentation, depth_colormap, dist_image, plot_fig

@spaces.GPU(duration=30)
def test_process_img(image):
    image = utils.resize(image)
    image_segmentation, objects_data = img_seg.predict(image)
    depthmap, depth_colormap = depth_estimator.make_prediction(image)
    return image_segmentation, objects_data, depthmap, depth_colormap

@spaces.GPU(duration=60)  # Longer duration for video processing
def process_video(vid_path=None):
    vid_cap = cv2.VideoCapture(vid_path)
    while vid_cap.isOpened():
        ret, frame = vid_cap.read()        
        if ret:
            print("making predictions ....")
            frame = utils.resize(frame)
            image_segmentation, objects_data = img_seg.predict(frame)
            depthmap, depth_colormap = depth_estimator.make_prediction(frame)
            dist_image = utils.draw_depth_info(frame, depthmap, objects_data)
            yield cv2.cvtColor(image_segmentation, cv2.COLOR_BGR2RGB), depth_colormap, cv2.cvtColor(dist_image, cv2.COLOR_BGR2RGB)
    
    return None

def update_segmentation_options(options):
    img_seg.is_show_bounding_boxes = True if 'Show Boundary Box' in options else False
    img_seg.is_show_segmentation = True if 'Show Segmentation Region' in options else False
    img_seg.is_show_segmentation_boundary = True if 'Show Segmentation Boundary' in options else False

def update_confidence_threshold(thres_val):
    img_seg.confidence_threshold = thres_val/100

@spaces.GPU(duration=10)  # Short duration for model loading
def model_selector(model_type):
    global img_seg, depth_estimator
    
    if "Small - Better performance and less accuracy" == model_type:
        midas_model, yolo_model = "midas_v21_small_256", "yolov8s-seg"
    elif "Medium - Balanced performance and accuracy" == model_type:
        midas_model, yolo_model = "dpt_hybrid_384", "yolov8m-seg"
    elif "Large - Slow performance and high accuracy" == model_type:
        midas_model, yolo_model = "dpt_large_384", "yolov8l-seg"
    else:
        midas_model, yolo_model = "midas_v21_small_256", "yolov8s-seg"

    img_seg = ImageSegmenter(model_type=yolo_model)
    depth_estimator = MonocularDepthEstimator(model_type=midas_model)

def cancel():
    global CANCEL_PROCESSING
    CANCEL_PROCESSING = True

if __name__ == "__main__":
    # gradio gui app
    with gr.Blocks() as my_app:
        # title
        gr.Markdown("<h1><center>Simultaneous Segmentation and Depth Estimation</center></h1>")
        gr.Markdown("<h3><center>Created by Vaishanth</center></h3>")
        gr.Markdown("<h3><center>This model estimates the depth of segmented objects.</center></h3>")

        # tabs
        with gr.Tab("Image"):
            with gr.Row():
                with gr.Column(scale=1):
                    img_input = gr.Image()
                    model_type_img = gr.Dropdown(
                        ["Small - Better performance and less accuracy", 
                         "Medium - Balanced performance and accuracy", 
                         "Large - Slow performance and high accuracy"], 
                        label="Model Type", value="Small - Better performance and less accuracy",
                        info="Select the inference model before running predictions!")
                    options_checkbox_img = gr.CheckboxGroup(["Show Boundary Box", "Show Segmentation Region", "Show Segmentation Boundary"], label="Options")
                    conf_thres_img = gr.Slider(1, 100, value=60, label="Confidence Threshold", info="Choose the threshold above which objects should be detected")
                    submit_btn_img = gr.Button(value="Predict")                    

                with gr.Column(scale=2):
                    with gr.Row():
                        segmentation_img_output = gr.Image(height=300, label="Segmentation")
                        depth_img_output = gr.Image(height=300, label="Depth Estimation")
                    
                    with gr.Row():
                        dist_img_output = gr.Image(height=300, label="Distance")
                        pcd_img_output = gr.Plot(label="Point Cloud")
            
            gr.Markdown("## Sample Images")
            gr.Examples(
                examples=[os.path.join(os.path.dirname(__file__), "assets/images/baggage_claim.jpg"),
                          os.path.join(os.path.dirname(__file__), "assets/images/kitchen_2.png"),
                          os.path.join(os.path.dirname(__file__), "assets/images/soccer.jpg"),
                          os.path.join(os.path.dirname(__file__), "assets/images/room_2.png"),
                          os.path.join(os.path.dirname(__file__), "assets/images/living_room.jpg")],
                inputs=img_input,
                outputs=[segmentation_img_output, depth_img_output, dist_img_output, pcd_img_output],
                fn=process_image,
                cache_examples=True,
            )

        with gr.Tab("Video"):
            with gr.Row():
                with gr.Column(scale=1):
                    vid_input = gr.Video()
                    model_type_vid = gr.Dropdown(
                        ["Small - Better performance and less accuracy", 
                         "Medium - Balanced performance and accuracy", 
                         "Large - Slow performance and high accuracy"], 
                        label="Model Type", value="Small - Better performance and less accuracy",
                        info="Select the inference model before running predictions!")
                    
                    options_checkbox_vid = gr.CheckboxGroup(["Show Boundary Box", "Show Segmentation Region", "Show Segmentation Boundary"], label="Options")
                    conf_thres_vid = gr.Slider(1, 100, value=60, label="Confidence Threshold", info="Choose the threshold above which objects should be detected")
                    with gr.Row():
                        cancel_btn = gr.Button(value="Cancel")
                        submit_btn_vid = gr.Button(value="Predict")
            
                with gr.Column(scale=2):
                    with gr.Row():
                        segmentation_vid_output = gr.Image(height=300, label="Segmentation")
                        depth_vid_output = gr.Image(height=300, label="Depth Estimation")
                    
                    with gr.Row():
                        dist_vid_output = gr.Image(height=300, label="Distance")
            
            gr.Markdown("## Sample Videos")
            gr.Examples(
                examples=[os.path.join(os.path.dirname(__file__), "assets/videos/input_video.mp4"),
                          os.path.join(os.path.dirname(__file__), "assets/videos/driving.mp4"),
                          os.path.join(os.path.dirname(__file__), "assets/videos/overpass.mp4"),
                          os.path.join(os.path.dirname(__file__), "assets/videos/walking.mp4")],
                inputs=vid_input,
            )

        # image tab logic
        submit_btn_img.click(process_image, inputs=img_input, outputs=[segmentation_img_output, depth_img_output, dist_img_output, pcd_img_output])
        options_checkbox_img.change(update_segmentation_options, options_checkbox_img, [])
        conf_thres_img.change(update_confidence_threshold, conf_thres_img, [])
        model_type_img.change(model_selector, model_type_img, [])

        # video tab logic
        submit_btn_vid.click(process_video, inputs=vid_input, outputs=[segmentation_vid_output, depth_vid_output, dist_vid_output])
        model_type_vid.change(model_selector, model_type_vid, [])
        cancel_btn.click(cancel, inputs=[], outputs=[])
        options_checkbox_vid.change(update_segmentation_options, options_checkbox_vid, [])
        conf_thres_vid.change(update_confidence_threshold, conf_thres_vid, [])       

    # Launch with appropriate queue settings for ZeroGPU
    my_app.queue(concurrency_count=1, max_size=10).launch()