app.py

from flask import Flask, request, jsonify, render_template
from detectron2.config import get_cfg
from detectron2.engine import DefaultPredictor
from detectron2.data import MetadataCatalog
from detectron2.utils.visualizer import Visualizer, ColorMode
import numpy as np
from PIL import Image
import io
import os
import requests
import gdown
from skimage import io
from torchvision.ops import box_iou
import torch

# Initialize Flask app
app = Flask(__name__)
cfg = None
# Google Drive file URL
GDRIVE_MODEL_URL = "https://drive.google.com/uc?id=18aEDo-kWOBhg8mAhnbpFkuM6bmmrBH4E"  # Replace 'your-file-id' with the actual file ID from Google Drive
LOCAL_MODEL_PATH = "model_final.pth"


def download_file_from_google_drive(id, destination):
    gdown.download(GDRIVE_MODEL_URL, LOCAL_MODEL_PATH, quiet=False)


file_id = "18aEDo-kWOBhg8mAhnbpFkuM6bmmrBH4E"
destination = "model_final.pth"
download_file_from_google_drive(file_id, destination)


# Download model from Google Drive if not already present locally
def download_model():
    if not os.path.exists(LOCAL_MODEL_PATH):
        response = requests.get(GDRIVE_MODEL_URL, stream=True)
        if response.status_code == 200:
            with open(LOCAL_MODEL_PATH, "wb") as f:
                f.write(response.content)
        else:
            raise Exception(
                f"Failed to download model from Google Drive: {response.status_code}"
            )


# Configuration and model setup
def setup_model(model_path):
    global cfg
    cfg = get_cfg()
    cfg.merge_from_file("config.yaml")  # Update with the config file path
    cfg.MODEL.WEIGHTS = model_path
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
    cfg.MODEL.DEVICE = "cpu"  # Use "cuda" for GPU
    return DefaultPredictor(cfg)


# Ensure model is available
predictor = setup_model(LOCAL_MODEL_PATH)

# Define expected parts and costs
expected_parts = ["headlamp", "rear_bumper", "door", "hood", "front_bumper"]
cost_dict = {
    "headlamp": 300,
    "rear_bumper": 250,
    "door": 200,
    "hood": 220,
    "front_bumper": 250,
    "other": 150,
}


@app.route("/")
def home():
    return render_template("index.html")


@app.route("/upload", methods=["POST"])
def upload():
    if "file" not in request.files:
        return jsonify({"error": "No file uploaded"}), 400

    file = request.files["file"]
    if file.filename == "":
        return jsonify({"error": "No file selected"}), 400

    # Load image
    image = io.imread(file)
    image_np = image

    # Run model prediction
    outputs = predictor(image_np)
    instances = outputs["instances"].to("cpu")
    class_names = MetadataCatalog.get(cfg.DATASETS.TEST[0]).thing_classes

    # Extract bounding boxes and class IDs
    boxes = instances.pred_boxes.tensor.numpy()
    class_ids = instances.pred_classes.numpy()

    # Filter overlapping boxes using IoU
    iou_threshold = 0.8
    keep_indices = []
    merged_boxes = set()

    for i in range(len(boxes)):
        if i in merged_boxes:
            continue
        keep_indices.append(i)
        for j in range(i + 1, len(boxes)):
            if j in merged_boxes:
                continue
            iou = box_iou(
                torch.tensor(boxes[i]).unsqueeze(0), torch.tensor(boxes[j]).unsqueeze(0)
            ).item()
            if iou > iou_threshold:
                merged_boxes.add(j)

    # Calculate total cost based on non-overlapping boxes
    total_cost = 0
    damage_details = []

    for idx in keep_indices:
        class_id = class_ids[idx]
        damaged_part = (
            class_names[class_id] if class_id < len(class_names) else "unknown"
        )
        if damaged_part not in expected_parts:
            damaged_part = "other"

        repair_cost = cost_dict.get(damaged_part, cost_dict["other"])
        total_cost += repair_cost

        damage_details.append({"part": damaged_part, "cost_usd": repair_cost})

    response = {"damages": damage_details, "total_cost": total_cost}

    return jsonify(response)


@app.route("/fetch-image", methods=["POST"])
def fetchImage():
    file = None
    if "url" in request.form:
        url = request.form["url"]
        response = requests.get(url)
        file = io.BytesIO(response.content)
    elif "file" in request.files:
        file = request.files["file"]

    # Load image
    image = io.imread(file)
    image_np = image

    # Run model prediction
    outputs = predictor(image_np)
    instances = outputs["instances"].to("cpu")
    class_names = MetadataCatalog.get(cfg.DATASETS.TEST[0]).thing_classes

    # Extract bounding boxes and class IDs
    boxes = instances.pred_boxes.tensor.numpy()
    class_ids = instances.pred_classes.numpy()

    # Filter overlapping boxes using IoU
    iou_threshold = 0.8
    keep_indices = []
    merged_boxes = set()

    for i in range(len(boxes)):
        if i in merged_boxes:
            continue
        keep_indices.append(i)
        for j in range(i + 1, len(boxes)):
            if j in merged_boxes:
                continue
            iou = box_iou(
                torch.tensor(boxes[i]).unsqueeze(0), torch.tensor(boxes[j]).unsqueeze(0)
            ).item()
            if iou > iou_threshold:
                merged_boxes.add(j)

    # Calculate total cost based on non-overlapping boxes
    total_cost = 0
    damage_details = []

    for idx in keep_indices:
        class_id = class_ids[idx]
        damaged_part = (
            class_names[class_id] if class_id < len(class_names) else "unknown"
        )
        if damaged_part not in expected_parts:
            damaged_part = "other"

        repair_cost = cost_dict.get(damaged_part, cost_dict["other"])
        total_cost += repair_cost

        damage_details.append({"part": damaged_part, "cost_usd": repair_cost})

    response = {"damages": damage_details, "total_cost": total_cost}

    return jsonify(response)


if __name__ == "__main__":
    app.run(host="0.0.0.0", port=7860)