app.py

import streamlit as st
import sahi.utils.file
import sahi.utils.mmdet
from sahi import AutoDetectionModel
from PIL import Image
import random
from utils import sahi_mmdet_inference
from streamlit_image_comparison import image_comparison

MMDET_YOLOX_TINY_MODEL_URL = "https://huggingface.co/fcakyon/mmdet-yolox-tiny/resolve/main/yolox_tiny_8x8_300e_coco_20211124_171234-b4047906.pth"
MMDET_YOLOX_TINY_MODEL_PATH = "yolox.pt"
MMDET_YOLOX_TINY_CONFIG_URL = "https://huggingface.co/fcakyon/mmdet-yolox-tiny/raw/main/yolox_tiny_8x8_300e_coco.py"
MMDET_YOLOX_TINY_CONFIG_PATH = "config.py"

IMAGE_TO_URL = {
    "apple_tree.jpg": "https://user-images.githubusercontent.com/34196005/142730935-2ace3999-a47b-49bb-83e0-2bdd509f1c90.jpg",
    "highway.jpg": "https://user-images.githubusercontent.com/34196005/142730936-1b397756-52e5-43be-a949-42ec0134d5d8.jpg",
    "highway2.jpg": "https://user-images.githubusercontent.com/34196005/142742871-bf485f84-0355-43a3-be86-96b44e63c3a2.jpg",
    "highway3.jpg": "https://user-images.githubusercontent.com/34196005/142742872-1fefcc4d-d7e6-4c43-bbb7-6b5982f7e4ba.jpg",
    "highway2-yolox.jpg": "https://user-images.githubusercontent.com/34196005/143309873-c0c1f31c-c42e-4a36-834e-da0a2336bb19.jpg",
    "highway2-sahi.jpg": "https://user-images.githubusercontent.com/34196005/143309867-42841f5a-9181-4d22-b570-65f90f2da231.jpg",
}


@st.cache(allow_output_mutation=True, show_spinner=False)
def download_comparison_images():
    sahi.utils.file.download_from_url(
        "https://user-images.githubusercontent.com/34196005/143309873-c0c1f31c-c42e-4a36-834e-da0a2336bb19.jpg",
        "highway2-yolox.jpg",
    )
    sahi.utils.file.download_from_url(
        "https://user-images.githubusercontent.com/34196005/143309867-42841f5a-9181-4d22-b570-65f90f2da231.jpg",
        "highway2-sahi.jpg",
    )


@st.cache(allow_output_mutation=True, show_spinner=False)
def get_model():
    sahi.utils.file.download_from_url(
        MMDET_YOLOX_TINY_MODEL_URL,
        MMDET_YOLOX_TINY_MODEL_PATH,
    )
    sahi.utils.file.download_from_url(
        MMDET_YOLOX_TINY_CONFIG_URL,
        MMDET_YOLOX_TINY_CONFIG_PATH,
    )

    detection_model = AutoDetectionModel.from_pretrained(
        model_type='mmdet',
        model_path=MMDET_YOLOX_TINY_MODEL_PATH,
        config_path=MMDET_YOLOX_TINY_CONFIG_PATH,
        confidence_threshold=0.5,
        device="cpu",
    )
    return detection_model


class SpinnerTexts:
    def __init__(self):
        self.ind_history_list = []
        self.text_list = [
            "Meanwhile check out [MMDetection Colab notebook of SAHI](https://colab.research.google.com/github/obss/sahi/blob/main/demo/inference_for_mmdetection.ipynb)!",
            "Meanwhile check out [YOLOv5 Colab notebook of SAHI](https://colab.research.google.com/github/obss/sahi/blob/main/demo/inference_for_yolov5.ipynb)!",
            "Meanwhile check out [aerial object detection with SAHI](https://blog.ml6.eu/how-to-detect-small-objects-in-very-large-images-70234bab0f98?gi=b434299595d4)!",
            "Meanwhile check out [COCO Utilities of SAHI](https://github.com/obss/sahi/blob/main/docs/COCO.md)!",
            "Meanwhile check out [FiftyOne utilities of SAHI](https://github.com/obss/sahi#fiftyone-utilities)!",
            "Meanwhile [give a Github star to SAHI](https://github.com/obss/sahi/stargazers)!",
            "Meanwhile see [how easy is to install SAHI](https://github.com/obss/sahi#getting-started)!",
            "Meanwhile check out [Medium blogpost of SAHI](https://medium.com/codable/sahi-a-vision-library-for-performing-sliced-inference-on-large-images-small-objects-c8b086af3b80)!",
            "Meanwhile try out [YOLOv5 HF Spaces demo of SAHI](https://huggingface.co/spaces/fcakyon/sahi-yolov5)!",
        ]

    def _store(self, ind):
        if len(self.ind_history_list) == 6:
            self.ind_history_list.pop(0)
        self.ind_history_list.append(ind)

    def get(self):
        ind = 0
        while ind in self.ind_history_list:
            ind = random.randint(0, len(self.text_list) - 1)
        self._store(ind)
        return self.text_list[ind]


st.set_page_config(
    page_title="Small Object Detection with SAHI + YOLOX",
    page_icon="🚀",
    layout="centered",
    initial_sidebar_state="auto",
)

download_comparison_images()

if "last_spinner_texts" not in st.session_state:
    st.session_state["last_spinner_texts"] = SpinnerTexts()

if "output_1" not in st.session_state:
    st.session_state["output_1"] = Image.open("highway2-yolox.jpg")

if "output_2" not in st.session_state:
    st.session_state["output_2"] = Image.open("highway2-sahi.jpg")

st.markdown(
    """
    <h2 style='text-align: center'>
    Small Object Detection <br />
    with SAHI + YOLOX
    </h2>
    """,
    unsafe_allow_html=True,
)
st.markdown(
    """
    <p style='text-align: center'>
    <a href='https://github.com/obss/sahi' target='_blank'>SAHI Github</a> | <a href='https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox' target='_blank'>YOLOX Github</a> | <a href='https://huggingface.co/spaces/fcakyon/sahi-yolov5' target='_blank'>SAHI+YOLOv5 Demo</a>
    <br />
    Follow me for more! <a href='https://twitter.com/fcakyon' target='_blank'> <img src="https://img.icons8.com/color/48/000000/twitter--v1.png" height="30"></a><a href='https://github.com/fcakyon' target='_blank'><img src="https://img.icons8.com/fluency/48/000000/github.png" height="27"></a><a href='https://www.linkedin.com/in/fcakyon/' target='_blank'><img src="https://img.icons8.com/fluency/48/000000/linkedin.png" height="30"></a> <a href='https://fcakyon.medium.com/' target='_blank'><img src="https://img.icons8.com/ios-filled/48/000000/medium-monogram.png" height="26"></a>
    </p>
    """,
    unsafe_allow_html=True,
)

st.write("##")

with st.expander("Usage"):
    st.markdown(
        """
        <p>
        1. Upload or select the input image 🖼️
        <br />
        2. (Optional) Set SAHI parameters ✔️
        <br />
        3. Press to "🚀 Perform Prediction"
        <br />
        4. Enjoy sliding image comparison 🔥
        </p>
        """,
        unsafe_allow_html=True,
    )

st.write("##")

col1, col2, col3 = st.columns([6, 1, 6])
with col1:
    st.markdown(f"##### Set input image:")

    # set input image by upload
    image_file = st.file_uploader(
        "Upload an image to test:", type=["jpg", "jpeg", "png"]
    )

    # set input image from exapmles
    def slider_func(option):
        option_to_id = {
            "apple_tree.jpg": str(1),
            "highway.jpg": str(2),
            "highway2.jpg": str(3),
            "highway3.jpg": str(4),
        }
        return option_to_id[option]

    slider = st.select_slider(
        "Or select from example images:",
        options=["apple_tree.jpg", "highway.jpg", "highway2.jpg", "highway3.jpg"],
        format_func=slider_func,
        value="highway2.jpg",
    )

    # visualize input image
    if image_file is not None:
        image = Image.open(image_file)
    else:
        image = sahi.utils.cv.read_image_as_pil(IMAGE_TO_URL[slider])
    st.image(image, width=300)

with col3:
    st.markdown(f"##### Set SAHI parameters:")

    slice_size = st.number_input("slice_size", min_value=256, value=512, step=256)
    overlap_ratio = st.number_input(
        "overlap_ratio", min_value=0.0, max_value=0.6, value=0.2, step=0.2
    )
    postprocess_type = st.selectbox(
        "postprocess_type", options=["NMS", "GREEDYNMM"], index=0
    )
    postprocess_match_metric = st.selectbox(
        "postprocess_match_metric", options=["IOU", "IOS"], index=0
    )
    postprocess_match_threshold = st.number_input(
        "postprocess_match_threshold", value=0.5, step=0.1
    )
    postprocess_class_agnostic = st.checkbox("postprocess_class_agnostic", value=True)

col1, col2, col3 = st.columns([4, 3, 4])
with col2:
    submit = st.button("🚀 Perform Prediction")

if submit:
    # perform prediction
    with st.spinner(
        text="Downloading model weight.. "
        + st.session_state["last_spinner_texts"].get()
    ):
        detection_model = get_model()

    image_size = 416

    with st.spinner(
        text="Performing prediction.. " + st.session_state["last_spinner_texts"].get()
    ):
        output_1, output_2 = sahi_mmdet_inference(
            image,
            detection_model,
            image_size=image_size,
            slice_height=slice_size,
            slice_width=slice_size,
            overlap_height_ratio=overlap_ratio,
            overlap_width_ratio=overlap_ratio,
            postprocess_type=postprocess_type,
            postprocess_match_metric=postprocess_match_metric,
            postprocess_match_threshold=postprocess_match_threshold,
            postprocess_class_agnostic=postprocess_class_agnostic,
        )

    st.session_state["output_1"] = output_1
    st.session_state["output_2"] = output_2

st.markdown(f"##### YOLOX Standard vs SAHI Prediction:")
static_component = image_comparison(
    img1=st.session_state["output_1"],
    img2=st.session_state["output_2"],
    label1="YOLOX",
    label2="SAHI+YOLOX",
    width=700,
    starting_position=50,
    show_labels=True,
    make_responsive=True,
    in_memory=True,
)
st.markdown(
    """
    <p style='text-align: center'>
    prepared with <a href='https://github.com/fcakyon/streamlit-image-comparison' target='_blank'>streamlit-image-comparison</a>
    </p>
    """,
    unsafe_allow_html=True,
)