app.py

import streamlit as st
import sahi.utils.mmdet
import sahi.model
import sahi.predict
from PIL import Image
import numpy
import random


MMDET_YOLACT_MODEL_URL = "https://download.openmmlab.com/mmdetection/v2.0/yolact/yolact_r50_1x8_coco/yolact_r50_1x8_coco_20200908-f38d58df.pth"
MMDET_YOLOX_MODEL_URL = "https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_tiny_8x8_300e_coco/yolox_tiny_8x8_300e_coco_20210806_234250-4ff3b67e.pth"
MMDET_FASTERRCNN_MODEL_URL = "https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth"

# Images
sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142730935-2ace3999-a47b-49bb-83e0-2bdd509f1c90.jpg",
    "apple_tree.jpg",
)
sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142730936-1b397756-52e5-43be-a949-42ec0134d5d8.jpg",
    "highway.jpg",
)

sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142742871-bf485f84-0355-43a3-be86-96b44e63c3a2.jpg",
    "highway2.jpg",
)

sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142742872-1fefcc4d-d7e6-4c43-bbb7-6b5982f7e4ba.jpg",
    "highway3.jpg",
)


@st.cache(allow_output_mutation=True, show_spinner=False)
def get_mmdet_model(model_name: str):
    if model_name == "yolact":
        model_path = "yolact.pt"
        sahi.utils.file.download_from_url(
            MMDET_YOLACT_MODEL_URL,
            model_path,
        )
        config_path = sahi.utils.mmdet.download_mmdet_config(
            model_name="yolact", config_file_name="yolact_r50_1x8_coco.py"
        )
    elif model_name == "yolox":
        model_path = "yolox.pt"
        sahi.utils.file.download_from_url(
            MMDET_YOLOX_MODEL_URL,
            model_path,
        )
        config_path = sahi.utils.mmdet.download_mmdet_config(
            model_name="yolox", config_file_name="yolox_tiny_8x8_300e_coco.py"
        )
    elif model_name == "faster_rcnn":
        model_path = "faster_rcnn.pt"
        sahi.utils.file.download_from_url(
            MMDET_FASTERRCNN_MODEL_URL,
            model_path,
        )
        config_path = sahi.utils.mmdet.download_mmdet_config(
            model_name="faster_rcnn", config_file_name="faster_rcnn_r50_fpn_2x_coco.py"
        )

    detection_model = sahi.model.MmdetDetectionModel(
        model_path=model_path,
        config_path=config_path,
        confidence_threshold=0.4,
        device="cpu",
    )
    return detection_model


def sahi_mmdet_inference(
    image,
    detection_model,
    slice_height=512,
    slice_width=512,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
    image_size=640,
    postprocess_type="UNIONMERGE",
    postprocess_match_metric="IOS",
    postprocess_match_threshold=0.5,
    postprocess_class_agnostic=False,
):

    # standard inference
    prediction_result_1 = sahi.predict.get_prediction(
        image=image, detection_model=detection_model, image_size=image_size
    )
    visual_result_1 = sahi.utils.cv.visualize_object_predictions(
        image=numpy.array(image),
        object_prediction_list=prediction_result_1.object_prediction_list,
    )
    output_1 = Image.fromarray(visual_result_1["image"])

    # sliced inference
    prediction_result_2 = sahi.predict.get_sliced_prediction(
        image=image,
        detection_model=detection_model,
        image_size=image_size,
        slice_height=slice_height,
        slice_width=slice_width,
        overlap_height_ratio=overlap_height_ratio,
        overlap_width_ratio=overlap_width_ratio,
        postprocess_type=postprocess_type,
        postprocess_match_metric=postprocess_match_metric,
        postprocess_match_threshold=postprocess_match_threshold,
        postprocess_class_agnostic=postprocess_class_agnostic,
    )
    visual_result_2 = sahi.utils.cv.visualize_object_predictions(
        image=numpy.array(image),
        object_prediction_list=prediction_result_2.object_prediction_list,
    )

    output_2 = Image.fromarray(visual_result_2["image"])

    return output_1, output_2


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

st.markdown(
    """<h2 style='text-align: center'>
    Small Object Detection <br />
    with SAHI + MMDetection
    </h2>""",
    unsafe_allow_html=True,
)
st.markdown(
    "<p style='text-align: center'>SAHI is a lightweight vision library for performing large scale object detection and instance segmentation.. <a href='https://github.com/obss/sahi'>SAHI Github</a> | <a href='https://medium.com/codable/sahi-a-vision-library-for-performing-sliced-inference-on-large-images-small-objects-c8b086af3b80'>SAHI Blog</a> | <a href='https://huggingface.co/spaces/fcakyon/sahi-yolov5'>SAHI+YOLOv5 Demo</a> </p>",
    unsafe_allow_html=True,
)

st.markdown(
    "<h3 style='text-align: center'> Parameters: </h1>",
    unsafe_allow_html=True,
)
col1, col2, col3 = st.columns([6, 1, 6])
with col1:
    image_file = st.file_uploader(
        "Upload an image to test:", type=["jpg", "jpeg", "png"]
    )

    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,
    )
    image = Image.open(slider)
    st.image(image, caption=slider, width=300)
with col3:
    model_name = st.selectbox(
        "Select MMDetection model:", ("faster_rcnn", "yolact", "yolox"), index=2
    )
    slice_size = st.number_input("slice_size", 256, value=512, step=256)
    overlap_ratio = st.number_input("overlap_ratio", 0.0, 0.6, value=0.2, step=0.2)
    postprocess_type = st.selectbox(
        "postprocess_type", options=["NMS", "UNIONMERGE"], index=1
    )
    postprocess_match_metric = st.selectbox(
        "postprocess_match_metric", options=["IOU", "IOS"], index=1
    )
    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([6, 1, 6])
with col2:
    submit = st.button("Submit")

if image_file is not None:
    image = Image.open(image_file)
else:
    image = Image.open(slider)


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]


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

if submit:
    # perform prediction
    st.markdown(
        "<h3 style='text-align: center'> Results: </h1>",
        unsafe_allow_html=True,
    )

    with st.spinner(
        text="Downloading model weight.. "
        + st.session_state["last_spinner_texts"].get()
    ):
        detection_model = get_mmdet_model(model_name)

    if model_name == "yolox":
        image_size = 416
    else:
        image_size = 640

    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.markdown(f"##### Standard {model_name} Prediction:")
    st.image(output_1, width=700)
    st.markdown(f"##### Sliced {model_name} Prediction:")
    st.image(output_2, width=700)