app.py

import os
import cv2
import tqdm
import uuid
import logging

import torch
import spaces
import numpy as np
import gradio as gr
import imageio.v3 as iio
import supervision as sv

from pathlib import Path
from functools import lru_cache
from typing import List, Optional, Tuple

from PIL import Image
from transformers import AutoModelForObjectDetection, AutoImageProcessor
from transformers.image_utils import load_image


# Configuration constants
CHECKPOINTS = [
    "ustc-community/dfine_m_obj2coco",
    "ustc-community/dfine_m_obj365",
    "ustc-community/dfine_n_coco",
    "ustc-community/dfine_s_coco",
    "ustc-community/dfine_m_coco",
    "ustc-community/dfine_l_coco",
    "ustc-community/dfine_x_coco",
    "ustc-community/dfine_s_obj365",
    "ustc-community/dfine_l_obj365",
    "ustc-community/dfine_x_obj365",
    "ustc-community/dfine_s_obj2coco",
    "ustc-community/dfine_l_obj2coco_e25",
    "ustc-community/dfine_x_obj2coco",
]
DEFAULT_CHECKPOINT = CHECKPOINTS[0]
DEFAULT_CONFIDENCE_THRESHOLD = 0.3

TORCH_DTYPE = torch.float32

# Image
IMAGE_EXAMPLES = [
    {"path": "./examples/images/crossroad.jpg", "use_url": False, "url": "", "label": "Local Image"},
    {
        "path": None,
        "use_url": True,
        "url": "https://live.staticflickr.com/65535/33021460783_1646d43c54_b.jpg",
        "label": "Flickr Image",
    },
]

# Video
MAX_NUM_FRAMES = 250
BATCH_SIZE = 4
ALLOWED_VIDEO_EXTENSIONS = {".mp4", ".avi", ".mov"}
VIDEO_OUTPUT_DIR = Path("static/videos")
VIDEO_OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
VIDEO_EXAMPLES = [
    {"path": "./examples/videos/traffic.mp4", "label": "Local Video"},
    {"path": "./examples/videos/fast_and_furious.mp4", "label": "Local Video"},
    {"path": "./examples/videos/break_dance.mp4", "label": "Local Video"},
]

logging.basicConfig(
    level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)


@lru_cache(maxsize=3)
def get_model_and_processor(checkpoint: str):
    model = AutoModelForObjectDetection.from_pretrained(checkpoint, torch_dtype=TORCH_DTYPE)
    image_processor = AutoImageProcessor.from_pretrained(checkpoint)
    return model, image_processor


@spaces.GPU(duration=20)
def detect_objects(
    checkpoint: str,
    images: List[np.ndarray] | np.ndarray, 
    confidence_threshold: float = DEFAULT_CONFIDENCE_THRESHOLD,
    target_size: Optional[Tuple[int, int]] = None,
    batch_size: int = BATCH_SIZE,
):

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model, image_processor = get_model_and_processor(checkpoint)
    model = model.to(device)

    if isinstance(images, np.ndarray) and images.ndim == 4:
        images = [x for x in images]  # split video array into list of images

    batches = [images[i:i + batch_size] for i in range(0, len(images), batch_size)]

    results = []
    for batch in tqdm.tqdm(batches, desc="Processing frames"):

        # preprocess images
        inputs = image_processor(images=batch, return_tensors="pt")
        inputs = inputs.to(device).to(TORCH_DTYPE)

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

        # postprocess outputs
        if target_size:
            target_sizes = [target_size] * len(batch)
        else:
            target_sizes = [(image.shape[0], image.shape[1]) for image in batch]

        batch_results = image_processor.post_process_object_detection(
            outputs, target_sizes=target_sizes, threshold=confidence_threshold
        )

        results.extend(batch_results)

    # move results to cpu
    for i, result in enumerate(results):
        results[i] = {k: v.cpu() for k, v in result.items()}
    
    return results, model.config.id2label


def process_image(
    checkpoint: str = DEFAULT_CHECKPOINT,
    image: Optional[Image.Image] = None,
    url: Optional[str] = None,
    confidence_threshold: float = DEFAULT_CONFIDENCE_THRESHOLD,
):

    if (image is None) ^ bool(url):
        raise ValueError(f"Either image or url must be provided, but not both.")

    if url:
        image = load_image(url)

    results, id2label = detect_objects(
        checkpoint=checkpoint,
        images=[np.array(image)],
        confidence_threshold=confidence_threshold,
    )
    result = results[0] # first image in batch (we have batch size 1)

    annotations = []
    for label, score, box in zip(result["labels"], result["scores"], result["boxes"]):
        text_label = id2label[label.item()]
        formatted_label = f"{text_label} ({score:.2f})"
        x_min, y_min, x_max, y_max = box.cpu().numpy().round().astype(int)
        x_min = max(0, x_min)
        y_min = max(0, y_min)
        x_max = min(image.width - 1, x_max)
        y_max = min(image.height - 1, y_max)
        annotations.append(((x_min, y_min, x_max, y_max), formatted_label))

    return (image, annotations)


def get_target_size(image_height, image_width, max_size: int):
    if image_height < max_size and image_width < max_size:
        new_height, new_width = image_width, image_height
    elif image_height > image_width:
        new_height = max_size
        new_width = int(image_width * max_size / image_height)
    else:
        new_width = max_size
        new_height = int(image_height * max_size / image_width)
    
    # make even (for video codec compatibility)
    new_height = new_height // 2 * 2
    new_width = new_width // 2 * 2

    return new_width, new_height


def read_video_k_frames(video_path: str, k: int, read_every_i_frame: int = 1):
    cap = cv2.VideoCapture(video_path)
    frames = []
    i = 0
    progress_bar = tqdm.tqdm(total=k, desc="Reading frames")
    while cap.isOpened() and len(frames) < k:
        ret, frame = cap.read()
        if not ret:
            break
        if i % read_every_i_frame == 0:
            frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
            progress_bar.update(1)
        i += 1
    cap.release()
    progress_bar.close()
    return frames


def process_video(
    video_path: str,
    checkpoint: str,
    confidence_threshold: float = DEFAULT_CONFIDENCE_THRESHOLD,
    progress: gr.Progress = gr.Progress(track_tqdm=True),
) -> str:
    
    if not video_path or not os.path.isfile(video_path):
        raise ValueError(f"Invalid video path: {video_path}")

    ext = os.path.splitext(video_path)[1].lower()
    if ext not in ALLOWED_VIDEO_EXTENSIONS:
        raise ValueError(f"Unsupported video format: {ext}, supported formats: {ALLOWED_VIDEO_EXTENSIONS}")

    video_info = sv.VideoInfo.from_video_path(video_path)
    read_each_i_frame = video_info.fps // 25
    target_fps = video_info.fps / read_each_i_frame
    target_width, target_height = get_target_size(video_info.height, video_info.width, 1080)

    n_frames_to_read = min(MAX_NUM_FRAMES, video_info.total_frames // read_each_i_frame)
    frames = read_video_k_frames(video_path, n_frames_to_read, read_each_i_frame)
    frames = [cv2.resize(frame, (target_width, target_height), interpolation=cv2.INTER_CUBIC) for frame in frames]

    box_annotator = sv.BoxAnnotator(thickness=1)
    label_annotator = sv.LabelAnnotator(text_scale=0.5)

    results, id2label = detect_objects(
        images=np.array(frames),
        checkpoint=checkpoint,
        confidence_threshold=confidence_threshold,
        target_size=(target_height, target_width),
    )

    annotated_frames = []
    for frame, result in tqdm.tqdm(zip(frames, results), desc="Annotating frames", total=len(frames)):
        detections = sv.Detections.from_transformers(result, id2label=id2label)
        detections = detections.with_nms(threshold=0.95, class_agnostic=True)
        annotated_frame = box_annotator.annotate(scene=frame, detections=detections)
        annotated_frame = label_annotator.annotate(scene=annotated_frame, detections=detections)
        annotated_frames.append(annotated_frame)

    output_filename = os.path.join(VIDEO_OUTPUT_DIR, f"output_{uuid.uuid4()}.mp4")
    iio.imwrite(output_filename, annotated_frames, fps=target_fps, codec="h264")
    return output_filename



def create_image_inputs() -> List[gr.components.Component]:
    return [
        gr.Image(
            label="Upload Image",
            type="pil",
            sources=["upload", "webcam"],
            interactive=True,
            elem_classes="input-component",
        ),
        gr.Checkbox(label="Use Image URL Instead", value=False),
        gr.Textbox(
            label="Image URL",
            placeholder="https://example.com/image.jpg",
            visible=False,
            elem_classes="input-component",
        ),
        gr.Dropdown(
            choices=CHECKPOINTS,
            label="Select Model Checkpoint",
            value=DEFAULT_CHECKPOINT,
            elem_classes="input-component",
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=DEFAULT_CONFIDENCE_THRESHOLD,
            step=0.1,
            label="Confidence Threshold",
            elem_classes="input-component",
        ),
    ]


def create_video_inputs() -> List[gr.components.Component]:
    return [
        gr.Video(
            label="Upload Video",
            sources=["upload"],
            interactive=True,
            format="mp4",  # Ensure MP4 format
            elem_classes="input-component",
        ),
        gr.Dropdown(
            choices=CHECKPOINTS,
            label="Select Model Checkpoint",
            value=DEFAULT_CHECKPOINT,
            elem_classes="input-component",
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=DEFAULT_CONFIDENCE_THRESHOLD,
            step=0.1,
            label="Confidence Threshold",
            elem_classes="input-component",
        ),
    ]


def create_button_row() -> List[gr.Button]:
    return [
        gr.Button(
            f"Detect Objects", variant="primary", elem_classes="action-button"
        ),
        gr.Button(f"Clear", variant="secondary", elem_classes="action-button"),
    ]


# Gradio interface
with gr.Blocks(theme=gr.themes.Ocean()) as demo:
    gr.Markdown(
        """
        # Object Detection Demo
        Experience state-of-the-art object detection with USTC's D-Fine models.
         - **Image** and **Video** modes are supported.
         - Select a model and adjust the confidence threshold to see detections!
        """,
        elem_classes="header-text",
    )

    with gr.Tabs():
        with gr.Tab("Image"):
            with gr.Row():
                with gr.Column(scale=1, min_width=300):
                    with gr.Group():
                        (
                            image_input,
                            use_url,
                            url_input,
                            image_model_checkpoint,
                            image_confidence_threshold,
                        ) = create_image_inputs()
                        image_detect_button, image_clear_button = create_button_row()
                with gr.Column(scale=2):
                    image_output = gr.AnnotatedImage(
                        label="Detection Results",
                        show_label=True,
                        color_map=None,
                        elem_classes="output-component",
                    )
            gr.Examples(
                examples=[
                    [
                        example["path"],
                        example["use_url"],
                        example["url"],
                        DEFAULT_CHECKPOINT,
                        DEFAULT_CONFIDENCE_THRESHOLD,
                    ]
                    for example in IMAGE_EXAMPLES
                ],
                inputs=[
                    image_input,
                    use_url,
                    url_input,
                    image_model_checkpoint,
                    image_confidence_threshold,
                ],
                outputs=[image_output],
                fn=process_image,
                cache_examples=False,
                label="Select an image example to populate inputs",
            )

        with gr.Tab("Video"):
            gr.Markdown(
                f"The input video will be processed in ~25 FPS (up to {MAX_NUM_FRAMES} frames in result)."
            )
            with gr.Row():
                with gr.Column(scale=1, min_width=300):
                    with gr.Group():
                        video_input, video_checkpoint, video_confidence_threshold = create_video_inputs()
                        video_detect_button, video_clear_button = create_button_row()
                with gr.Column(scale=2):
                    video_output = gr.Video(
                        label="Detection Results",
                        format="mp4",  # Explicit MP4 format
                        elem_classes="output-component",
                    )

            gr.Examples(
                examples=[
                    [example["path"], DEFAULT_CHECKPOINT, DEFAULT_CONFIDENCE_THRESHOLD]
                    for example in VIDEO_EXAMPLES
                ],
                inputs=[video_input, video_checkpoint, video_confidence_threshold],
                outputs=[video_output],
                fn=process_video,
                cache_examples=False,
                label="Select a video example to populate inputs",
            )

    # Dynamic visibility for URL input
    use_url.change(
        fn=lambda x: gr.update(visible=x),
        inputs=use_url,
        outputs=url_input,
    )

    # Image clear button
    image_clear_button.click(
        fn=lambda: (
            None,
            False,
            "",
            DEFAULT_CHECKPOINT,
            DEFAULT_CONFIDENCE_THRESHOLD,
            None,
        ),
        outputs=[
            image_input,
            use_url,
            url_input,
            image_model_checkpoint,
            image_confidence_threshold,
            image_output,
        ],
    )

    # Video clear button
    video_clear_button.click(
        fn=lambda: (
            None,
            DEFAULT_CHECKPOINT,
            DEFAULT_CONFIDENCE_THRESHOLD,
            None,
        ),
        outputs=[
            video_input,
            video_checkpoint,
            video_confidence_threshold,
            video_output,
        ],
    )

    # Image detect button
    image_detect_button.click(
        fn=process_image,
        inputs=[
            image_model_checkpoint,
            image_input,
            url_input,
            image_confidence_threshold,
        ],
        outputs=[image_output],
    )

    # Video detect button
    video_detect_button.click(
        fn=process_video,
        inputs=[video_input, video_checkpoint, video_confidence_threshold],
        outputs=[video_output],
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch()