app.py

import datetime
import io
import time

import torch
import gradio as gr
import cv2

from transformers import AutoFeatureExtractor, AutoModelForObjectDetection

extractor = AutoFeatureExtractor.from_pretrained("hustvl/yolos-tiny")
model = AutoModelForObjectDetection.from_pretrained("hustvl/yolos-tiny")

BBOX_COLOR = [255, 0, 0]
PRED_THRESHOLD = 0.90

def composite_predictions(img, processed_predictions, show_video=False):
    interested_labels = processed_predictions["labels"] == 1 # only interested in people 
    scores = processed_predictions["scores"][interested_labels].tolist()
    boxes = [[int(j) for j in x] for x in processed_predictions["boxes"][interested_labels].tolist()]
    labels = [model.config.id2label[x] for x in processed_predictions["labels"][interested_labels].tolist()]
    
    for score, box, label in zip(scores, boxes, labels):
        cv2.rectangle(img, box, BBOX_COLOR, 1)
        cv2.putText(img, f"{label}: {score:0.2f}", (box[0]+2, box[1]+10), cv2.FONT_HERSHEY_SIMPLEX, 0.33, BBOX_COLOR, 1, cv2.LINE_AA)
    return img, len(boxes), datetime.datetime.now()

def process(img):
    inputs = extractor(images=img, return_tensors="pt")
    outputs = model(**inputs)
    h, w, _ = img.shape
    img_size = torch.tensor([(h, w)])
    processed = extractor.post_process_object_detection(outputs, PRED_THRESHOLD, img_size)
    
    # Composite image and prediction bounding boxes + labels prediction
    return composite_predictions(img, processed[0])


with gr.Blocks() as demo:
    stream = gr.State()
    with gr.Row():
        with gr.Column(scale=1, min_width=600):
            last_refresh_box = gr.Textbox(label="Last updated")
            attendance_label = gr.Label(label="Current Attendance")
        with gr.Row():
            with gr.Column(scale=1, min_width=600):
                webcam = gr.Webcam(streaming=True)
                output = gr.Image(label="Composite", visible=True)
    webcam.stream(process, [webcam], [output, attendance_label, last_refresh_box])

demo.queue().launch()