app-fix-dino.py

import gradio as gr
from dotenv import load_dotenv
from roboflow import Roboflow
import tempfile
import os
import requests
import cv2
import numpy as np
from dds_cloudapi_sdk import Config, Client
from dds_cloudapi_sdk.tasks.dinox import DinoxTask
from dds_cloudapi_sdk.tasks.types import DetectionTarget
from dds_cloudapi_sdk import TextPrompt
import subprocess

# ========== Konfigurasi ========== 
load_dotenv()

# Roboflow Config
rf_api_key = os.getenv("ROBOFLOW_API_KEY")
workspace = os.getenv("ROBOFLOW_WORKSPACE")
project_name = os.getenv("ROBOFLOW_PROJECT")
model_version = int(os.getenv("ROBOFLOW_MODEL_VERSION"))

# DINO-X Config
DINOX_API_KEY = os.getenv("DINO_X_API_KEY")
DINOX_PROMPT = "beverage . bottle . cans . mixed box"  # Customize sesuai produk kompetitor : food . drink

# Inisialisasi Model
rf = Roboflow(api_key=rf_api_key)
project = rf.workspace(workspace).project(project_name)
yolo_model = project.version(model_version).model

dinox_config = Config(DINOX_API_KEY)
dinox_client = Client(dinox_config)

# ========== Fungsi Deteksi Kombinasi ========== 
def detect_combined(image):
    with tempfile.NamedTemporaryFile(delete=False, suffix=".jpg") as temp_file:
        image.save(temp_file, format="JPEG")
        temp_path = temp_file.name

    try:
        # ========== [1] YOLO: Deteksi Produk Nestlé (Per Class) ========== 
        yolo_pred = yolo_model.predict(temp_path, confidence=50, overlap=80).json()

        # Hitung per class Nestlé
        nestle_class_count = {}
        nestle_boxes = []
        for pred in yolo_pred['predictions']:
            class_name = pred['class']
            nestle_class_count[class_name] = nestle_class_count.get(class_name, 0) + 1
            nestle_boxes.append((pred['x'], pred['y'], pred['width'], pred['height']))

        total_nestle = sum(nestle_class_count.values())

        # ========== [2] DINO-X: Deteksi Kompetitor ========== 
        image_url = dinox_client.upload_file(temp_path)
        task = DinoxTask(
            image_url=image_url,
            prompts=[TextPrompt(text=DINOX_PROMPT)],
            bbox_threshold=0.4,
            targets=[DetectionTarget.BBox]
        )
        dinox_client.run_task(task)
        dinox_pred = task.result.objects

        # Filter & Hitung Kompetitor
        competitor_class_count = {}
        competitor_boxes = []
        for obj in dinox_pred:
            dinox_box = obj.bbox
            # Filter objek yang sudah terdeteksi oleh YOLO (Overlap detection)
            if not is_overlap(dinox_box, nestle_boxes):  # Ignore if overlap with YOLO detections
                class_name = obj.category.strip().lower()  # Normalisasi nama kelas
                competitor_class_count[class_name] = competitor_class_count.get(class_name, 0) + 1
                competitor_boxes.append({
                    "class": class_name,
                    "box": dinox_box,
                    "confidence": obj.score
                })

        total_competitor = sum(competitor_class_count.values())

        # ========== [3] Format Output ========== 
        result_text = "Product Nestle\n\n"
        for class_name, count in nestle_class_count.items():
            result_text += f"{class_name}: {count}\n"
        result_text += f"\nTotal Products Nestle: {total_nestle}\n\n"

        #result_text += "Competitor Products\n\n"
        if competitor_class_count:
            result_text += f"Total Unclassified Products: {total_competitor}\n"  # Hanya total, tidak per kelas
        else:
            result_text += "No Unclassified Products detected\n"

        # ========== [4] Visualisasi ========== 
        img = cv2.imread(temp_path)

        # Nestlé (Hijau)
        for pred in yolo_pred['predictions']:
            x, y, w, h = pred['x'], pred['y'], pred['width'], pred['height']
            cv2.rectangle(img, (int(x-w/2), int(y-h/2)), (int(x+w/2), int(y+h/2)), (0,255,0), 2)
            cv2.putText(img, pred['class'], (int(x-w/2), int(y-h/2-10)), 
                       cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0,255,0), 3)

        # Kompetitor (Merah) dengan nama 'unclassified'
        for comp in competitor_boxes:
            x1, y1, x2, y2 = comp['box']
            
            # Define a list of target classes to rename
            unclassified_classes = ["beverage", "cans", "bottle", "mixed box"]
            
            # Normalize the class name to be case-insensitive and check if it's in the unclassified list
            display_name = "unclassified" if any(class_name in comp['class'].lower() for class_name in unclassified_classes) else comp['class']
            
            cv2.rectangle(img, (int(x1), int(y1)), (int(x2), int(y2)), (0, 0, 255), 2)
            cv2.putText(img, f"{display_name} {comp['confidence']:.2f}",
                        (int(x1), int(y1-10)), cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0, 0, 255), 3)
            
        output_path = "/tmp/combined_output.jpg"
        cv2.imwrite(output_path, img)

        return output_path, result_text

    except Exception as e:
        return temp_path, f"Error: {str(e)}"
    finally:
        os.remove(temp_path)

def is_overlap(box1, boxes2, threshold=0.3):
    # Fungsi untuk deteksi overlap bounding box
    x1_min, y1_min, x1_max, y1_max = box1
    for b2 in boxes2:
        x2, y2, w2, h2 = b2
        x2_min = x2 - w2/2
        x2_max = x2 + w2/2
        y2_min = y2 - h2/2
        y2_max = y2 + h2/2

        # Hitung area overlap
        dx = min(x1_max, x2_max) - max(x1_min, x2_min)
        dy = min(y1_max, y2_max) - max(y1_min, y2_min)
        if (dx >= 0) and (dy >= 0):
            area_overlap = dx * dy
            area_box1 = (x1_max - x1_min) * (y1_max - y1_min)
            if area_overlap / area_box1 > threshold:
                return True
    return False

# ========== Fungsi untuk Deteksi Video ========== 

def convert_video_to_mp4(input_path, output_path):
    try:
        subprocess.run(['ffmpeg', '-i', input_path, '-vcodec', 'libx264', '-acodec', 'aac', output_path], check=True)
        return output_path
    except subprocess.CalledProcessError as e:
        return None, f"Error converting video: {e}"

def detect_objects_in_video(video_path):
    temp_output_path = "/tmp/output_video.mp4"
    temp_frames_dir = tempfile.mkdtemp()
    frame_count = 0
    previous_detections = {}  # For storing previous frame's object detections

    try:
        # Convert video to MP4 if necessary
        if not video_path.endswith(".mp4"):
            video_path, err = convert_video_to_mp4(video_path, temp_output_path)
            if not video_path:
                return None, f"Video conversion error: {err}"

        # Read video and process frames
        video = cv2.VideoCapture(video_path)
        frame_rate = int(video.get(cv2.CAP_PROP_FPS))
        frame_width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
        frame_height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
        frame_size = (frame_width, frame_height)

        # VideoWriter for output video
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
        output_video = cv2.VideoWriter(temp_output_path, fourcc, frame_rate, frame_size)

        while True:
            ret, frame = video.read()
            if not ret:
                break

            # Save frame temporarily for predictions
            frame_path = os.path.join(temp_frames_dir, f"frame_{frame_count}.jpg")
            cv2.imwrite(frame_path, frame)

            # Process predictions for the current frame
            predictions = yolo_model.predict(frame_path, confidence=50, overlap=80).json()

            # Track current frame detections
            current_detections = {}
            for prediction in predictions['predictions']:
                class_name = prediction['class']
                x, y, w, h = prediction['x'], prediction['y'], prediction['width'], prediction['height']
                # Generate a unique ID for each detection based on the bounding box
                object_id = f"{class_name}_{x}_{y}_{w}_{h}"

                # Track each detected object individually
                if object_id not in current_detections:
                    current_detections[object_id] = class_name

                # Draw bounding box for detected objects
                cv2.rectangle(frame, (int(x-w/2), int(y-h/2)), (int(x+w/2), int(y+h/2)), (0,255,0), 2)
                cv2.putText(frame, class_name, (int(x-w/2), int(y-h/2-10)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 2)

            # Update counts for objects
            object_counts = {}
            for detection_id in current_detections.keys():
                class_name = current_detections[detection_id]
                object_counts[class_name] = object_counts.get(class_name, 0) + 1

            # Generate display text for counts
            count_text = ""
            total_product_count = 0
            for class_name, count in object_counts.items():
                count_text += f"{class_name}: {count}\n"
                total_product_count += count
            count_text += f"\nTotal Product: {total_product_count}"

            # Overlay the counts text onto the frame
            y_offset = 20
            for line in count_text.split("\n"):
                cv2.putText(frame, line, (10, y_offset), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 255), 2)
                y_offset += 30  # Move down for next line

            # Write processed frame to output video
            output_video.write(frame)
            frame_count += 1

            # Update previous_detections for the next frame
            previous_detections = current_detections

        video.release()
        output_video.release()

        return temp_output_path

    except Exception as e:
        return None, f"An error occurred: {e}"

# ========== Gradio Interface ========== 
with gr.Blocks(theme=gr.themes.Base(primary_hue="teal", secondary_hue="teal", neutral_hue="slate")) as iface:
    gr.Markdown("""<div style="text-align: center;"><h1>NESTLE - STOCK COUNTING</h1></div>""")
    
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(type="pil", label="Input Image")
            detect_image_button = gr.Button("Detect Image")
            output_image = gr.Image(label="Detect Object")
            output_text = gr.Textbox(label="Counting Object")
            detect_image_button.click(fn=detect_combined, inputs=input_image, outputs=[output_image, output_text])

        with gr.Column():
            input_video = gr.Video(label="Input Video")
            detect_video_button = gr.Button("Detect Video")
            output_video = gr.Video(label="Output Video")
            detect_video_button.click(fn=detect_objects_in_video, inputs=input_video, outputs=[output_video])

iface.launch()