app.py

import torch
import gradio as gr
from transformers import Owlv2Processor, Owlv2ForObjectDetection
import cv2
import spaces

# ===============================
# DEVICE
# ===============================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = Owlv2ForObjectDetection.from_pretrained(
    "google/owlv2-base-patch16-ensemble"
).to(device)

processor = Owlv2Processor.from_pretrained(
    "google/owlv2-base-patch16-ensemble"
)

# ===============================
# MAIN FUNCTION
# ===============================
@spaces.GPU
def query_image(img, text_queries, score_threshold):

    # Split queries (still required internally)
    text_queries = text_queries.split(",")

    # Prepare inputs
    inputs = processor(
        text=text_queries,
        images=img,
        return_tensors="pt"
    ).to(device)

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

    # Move outputs to CPU
    outputs.logits = outputs.logits.cpu()
    outputs.pred_boxes = outputs.pred_boxes.cpu()

    # Correct target size (IMPORTANT)
    target_sizes = torch.tensor([img.shape[:2]])

    # ✅ FIXED METHOD (important!)
    results = processor.post_process_grounded_object_detection(
        outputs=outputs,
        target_sizes=target_sizes
    )[0]

    boxes = results["boxes"]
    scores = results["scores"]

    output_boxes = []

    # Process detections
    for box, score in zip(boxes, scores):

        if score < score_threshold:
            continue

        x1, y1, x2, y2 = map(int, box.tolist())

        # Save ONLY coordinates
        output_boxes.append([x1, y1, x2, y2])

        # Draw rectangle ONLY (no labels)
        cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)

    return img, output_boxes


# ===============================
# GRADIO UI
# ===============================
demo = gr.Interface(
    fn=query_image,
    inputs=[
        gr.Image(type="numpy"),
        gr.Textbox(label="Classes (comma separated)"),
        gr.Slider(0, 1, value=0.1)
    ],
    outputs=[
        gr.Image(label="Bounding Boxes"),
        gr.JSON(label="Coordinates Only")
    ],
    title="OWLv2 Bounding Box Coordinates Only"
)

# Launch app
demo.launch()