app.py

from io import BytesIO
from collections import Counter
import requests
import torch
import re
import gradio as gr
from PIL import Image, ImageDraw
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection

# Load model once
model_id = "IDEA-Research/grounding-dino-tiny"
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(device)

# prompt normalization function
def normalize_prompt(text_prompt):
    # lowercase
    text_prompt = text_prompt.lower().strip()

    # replace common separators with "."
    text_prompt = re.sub(r"\s*(and|,|&)\s*", ".", text_prompt)

    # split words and remove empties
    parts = [p.strip() for p in text_prompt.split(".") if p.strip()]

    # rebuild as "cat. dog."
    return ". ".join(parts) + "."

def detect_objects(image_url, uploaded_image, text_prompt):
    try:
        # Load image
        if uploaded_image is not None:
            image = uploaded_image.convert("RGB")
        elif image_url:
            headers = {"User-Agent": "Mozilla/5.0"}
            response = requests.get(image_url, headers=headers, timeout=10)
            response.raise_for_status()
            image = Image.open(BytesIO(response.content)).convert("RGB")
        else:
            return None, "Please provide an image URL or upload an image."

        # Default prompt fallback
        display_prompt = text_prompt.strip() if text_prompt and text_prompt.strip() else "capsule"
        model_prompt = normalize_prompt(display_prompt)

        # Inference
        inputs = processor(images=image, text=model_prompt, return_tensors="pt").to(device)

        with torch.no_grad():
            outputs = model(**inputs)

        results = processor.post_process_grounded_object_detection(
            outputs,
            inputs.input_ids,
            threshold=0.4,
            target_sizes=[image.size[::-1]]
        )

        # Draw results
        draw = ImageDraw.Draw(image)
        detected_labels = []
        
        for result in results:
            boxes = result["boxes"]
            scores = result["scores"]
            labels = result["text_labels"]
        
            # Skip empty detections
            if len(boxes) == 0:
                continue
        
            for box, score, label in zip(boxes, scores, labels):
                box = box.tolist()
                detected_labels.append(label)
        
                x1, y1, x2, y2 = box
        
                # Draw bounding box
                draw.rectangle([x1, y1, x2, y2], outline="red", width=3)
        
                # Draw label
                draw.text((x1, max(0, y1 - 15)), f"{label} {score:.2f}", fill="red")

        if not detected_labels:
            searched_object = text_prompt.replace(".", ", ").strip(", ").strip()
        
            return f"""
                ## No {searched_object} found in the image
                
                Try:
                - lowering the threshold
                - using a clearer image
                - changing the detection prompt
                """, image

        counts = Counter(detected_labels)

        summary_rows = []
        
        for label, count in counts.items():
            summary_rows.append(
                f"<tr><td style='padding:4px 12px'>{label}</td>"
                f"<td style='padding:4px 12px'><b>{count}</b></td></tr>"
            )
        
        total_types = len(counts)
        
        summary = f"""
            <h3>Detected {total_types} object type(s) for: {display_prompt}</h3>
            
            <table style='border-collapse: collapse; width: 100%;'>
            <tr>
            <th style='text-align:left; padding:4px 12px;'>Object</th>
            <th style='text-align:left; padding:4px 12px;'>Count</th>
            </tr>
            {''.join(summary_rows)}
            </table>
            """

        return summary, image

    except Exception as e:
        return None, f"Error: {str(e)}"


app = gr.Interface(
    fn=detect_objects,
    inputs=[
        gr.Textbox(label="Image URL"),
        gr.Image(type="pil", label="Upload JPG/PNG"),
        gr.Textbox(label="Detection Prompt", placeholder="e.g. a cat")
    ],
    outputs=[
        gr.Markdown(label="Detection Summary"),
        gr.Image(label="Annotated Image")
    ],
    title="Grounding DINO Object Detection",
    description="Upload an image or provide an image URL, then enter objects to detect."
)

app.launch(server_name="0.0.0.0", server_port=7860)