app.py

import gradio as gr
import torch
import numpy as np
import cv2
from PIL import Image
import pandas as pd
from torchvision.transforms import Compose, Resize, ToTensor, Normalize
from segment_anything import SamPredictor, sam_model_registry
import os

# Load SAM and MiDaS models
def load_models():
    sam_checkpoint = "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth"
    if not os.path.exists(sam_checkpoint):
        raise FileNotFoundError("Please upload the SAM checkpoint file to the working directory.")

    device = "cuda" if torch.cuda.is_available() else "cpu"
    sam = sam_model_registry["vit_h"](checkpoint=sam_checkpoint).to(device)
    predictor = SamPredictor(sam)

    midas = torch.hub.load("intel-isl/MiDaS", "DPT_Large")
    midas.eval().to(device)
    midas_transform = Compose([
        Resize(384),
        ToTensor(),
        Normalize(mean=[0.5]*3, std=[0.5]*3)
    ])
    return predictor, midas, midas_transform

predictor, midas_model, midas_transform = load_models()

# Processing function
def process_image(image_pil):
    image_np = np.array(image_pil)
    img_h, img_w = image_np.shape[:2]

    # Real-world reference dimensions (adjust as needed)
    real_image_width_cm = 100
    real_image_height_cm = 75
    assumed_max_depth_cm = 100

    pixel_to_cm_x = real_image_width_cm / img_w
    pixel_to_cm_y = real_image_height_cm / img_h

    # SAM segmentation
    predictor.set_image(image_np)
    masks, _, _ = predictor.predict(multimask_output=False)

    # MiDaS depth estimation
    input_tensor = midas_transform(image_pil).unsqueeze(0).to(next(midas_model.parameters()).device)
    with torch.no_grad():
        depth_prediction = midas_model(input_tensor).squeeze().cpu().numpy()
    depth_resized = cv2.resize(depth_prediction, (img_w, img_h))

    # Object volume computation
    volume_data = []
    for i, mask in enumerate(masks):
        x, y, w, h = cv2.boundingRect(mask.astype(np.uint8))
        width_px = w
        height_px = h
        width_cm = width_px * pixel_to_cm_x
        height_cm = height_px * pixel_to_cm_y

        depth_masked = depth_resized[mask > 0.5]
        if depth_masked.size == 0:
            continue

        normalized_depth = (depth_masked - np.min(depth_resized)) / (np.max(depth_resized) - np.min(depth_resized) + 1e-6)
        depth_cm = np.mean(normalized_depth) * assumed_max_depth_cm
        volume_cm3 = round(depth_cm * width_cm * height_cm, 2)

        volume_data.append([
            f"Object #{i+1}",
            round(depth_cm, 2),
            round(width_cm, 2),
            round(height_cm, 2),
            volume_cm3
        ])

    if not volume_data:
        return image_pil, "No objects segmented."

    df = pd.DataFrame(volume_data, columns=["Object", "Length (Depth) cm", "Breadth (Width) cm", "Height cm", "Volume cm³"])
    return image_pil, df

# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# 📦 Volume Estimation using SAM + MiDaS")
    with gr.Row():
        image_input = gr.Image(type="pil", label="Upload Image")
        run_btn = gr.Button("Estimate Volume")
    with gr.Row():
        output_image = gr.Image(label="Original Image")
        volume_table = gr.Dataframe(headers=["Object", "Length (Depth) cm", "Breadth (Width) cm", "Height cm", "Volume cm³"])
    run_btn.click(fn=process_image, inputs=image_input, outputs=[output_image, volume_table])

demo.launch()