app.py

import gradio as gr
import onnxruntime as rt
import cv2
import numpy as np
from PIL import Image

H, W = 224, 224
classes=['aeroplane','bicycle','bird','boat','bottle','bus','car','cat','chair','cow','diningtable',
         'dog','horse','motorbike','person','pottedplant','sheep','sofa','train','tvmonitor']

providers = ['CPUExecutionProvider']

m = rt.InferenceSession("./model/yolo_efficient.onnx", providers=providers)

def nms(final_boxes, scores, IOU_threshold=0):
    scores = np.array(scores)
    final_boxes = np.array(final_boxes)

    boxes = final_boxes[...,:-1]

    boxes = [list(map(int, i)) for i in boxes]
    boxes = np.array(boxes)
    
    x1 = boxes[:, 0]
    y1 = boxes[:, 1]
    x2 = boxes[:, 2]
    y2 = boxes[:, 3]


    area = (x2 - x1)*(y2 - y1)

    order = np.argsort(scores)

    pick = []

    while len(order) > 0:
        last = len(order)-1
        i = order[last]
        pick.append(i)

        suppress = [last]

        if len(order)==0:
            break

        for pos in range(last):
            j = order[pos]

            xx1 = max(x1[i], x1[j])
            yy1 = max(y1[i], y1[j])
            xx2 = min(x2[i], x2[j])
            yy2 = min(y2[i], y2[j])

            w = max(0, xx2-xx1+1)
            h = max(0, yy2-yy1+1)

            overlap = float(w*h)/area[j]

            if overlap > IOU_threshold:
                suppress.append(pos)

        order = np.delete(order, suppress)

    return final_boxes[pick]

def detect_obj(input_image, obj_threshold, bb_threshold):
    try:
        
        image = np.array(input_image)
        image = cv2.resize(image, (H, W))
        img = image

        image = image.astype(np.float32)
        image = np.expand_dims(image, axis=0)

        output = m.run(['reshape'], {"input": image})
        output = np.squeeze(output, axis=0)

        object_positions = np.concatenate(
                [np.stack(np.where(output[..., 0]>=obj_threshold), axis=-1),
                 np.stack(np.where(output[..., 5]>=obj_threshold), axis=-1)], axis=0
        )

        selected_output = []
        for indices in object_positions:
                selected_output.append(output[indices[0]][indices[1]][indices[2]])
        selected_output = np.array(selected_output)

        final_boxes = []
        final_scores = []

        for i,pos in enumerate(object_positions):
            for j in range(2):
                if selected_output[i][j*5]>obj_threshold:
                    output_box = np.array(output[pos[0]][pos[1]][pos[2]][(j*5)+1:(j*5)+5], dtype=float)

                    x_centre = (np.array(pos[1], dtype=float) + output_box[0])*32
                    y_centre = (np.array(pos[2], dtype=float) + output_box[1])*32

                    x_width, y_height = abs(W*output_box[2]), abs(H*output_box[3])

                    x_min, y_min = int(x_centre - (x_width/2)), int(y_centre-(y_height/2))
                    x_max, y_max = int(x_centre+(x_width/2)), int(y_centre + (y_height/2))

                    if(x_min<0):x_min=0
                    if(y_min<0):y_min=0
                    if(x_max<0):x_max=0
                    if(y_max<0):y_max=0

                    final_boxes.append(
                        [x_min, y_min, x_max, y_max, str(classes[np.argmax(selected_output[..., 10:], axis=-1)[i]])]
                    )
                    final_scores.append(selected_output[i][j*5])

        final_boxes = np.array(final_boxes)

        nms_output = nms(final_boxes, final_scores, bb_threshold)

        for i in nms_output:
            cv2.rectangle(
                img,
                (int(i[0]), int(i[1])),
                (int(i[2]), int(i[3])), (255, 0, 0)
            )

            cv2.putText(
                img,
                i[-1],
                (int(i[0]), int(i[1])+15),
                cv2.FONT_HERSHEY_PLAIN, 1, (255, 0, 0), 1
            )

        output_pil_img = Image.fromarray(np.uint8(img)).convert('RGB')

        return output_pil_img

    except:
        return input_image



with gr.Blocks(title="YOLOS Object Detection - ClassCat", css=".gradio-container {background:lightyellow;}") as demo:
    gr.HTML('<h1>Yolo Object Detection</h1>')
    gr.HTML("<h4>supported objects are [aeroplane,bicycle,bird,boat,bottle,bus,car,cat,chair,cow,diningtable,dog,horse,motorbike,person,pottedplant,sheep,sofa,train,tvmonitor]</h4>")
    gr.HTML("<br>")
    with gr.Row():
        input_image = gr.Image(label="Input image", type="pil")
        output_image = gr.Image(label="Output image", type="pil")
    gr.HTML("<br>")
    gr.HTML("<h4>object centre detection threshold means the object centre will be considered a new object if it's value is above threshold</h4>")
    gr.HTML("<p>less means more objects</p>")
    gr.HTML("<h4>bounding box threshold is IOU value threshold. If intersection/union area of two bounding boxes are greater than threshold value the one box will be suppressed</h4>")
    gr.HTML("<p>more means more bounding boxes<p>")
    gr.HTML("<br>")

    obj_threshold = gr.Slider(0, 1.0, value=0.2, label=' object centre detection threshold')
    gr.HTML("<br>")
    bb_threshold = gr.Slider(0, 1.0, value=0.3, label=' bounding box draw threshold')
    gr.HTML("<br>")

    send_btn = gr.Button("Detect")
    gr.HTML("<br>")
    gr.Examples(['./samples/out_1.jpg'], inputs=input_image)

    send_btn.click(fn=detect_obj, inputs=[input_image, obj_threshold, bb_threshold], outputs=[output_image])



demo.launch(debug=True)