efficientdet_camera_inference

efficientdet_camera_inference.py

import os
import sys
import cv2
import time
import requests
import argparse
import tarfile
import numpy as np
import tensorflow.compat.v1 as tf

sys.path.append('.')
import inference

 
# min_score_thresh = 0.2
# max_boxes_to_draw = 100
# line_thickness = 4


def maybe_download(model_name):
    if os.path.exists(model_name):
        return

    url = "https://storage.googleapis.com/cloud-tpu-checkpoints/efficientdet/coco/{}.tar.gz".format(model_name)
    r = requests.get(url, allow_redirects=True)
    with open("{}.tar.gz".format(model_name), 'wb') as f:
        f.write(r.content)
    tf = tarfile.open("{}.tar.gz".format(model_name))
    tf.extractall()
    print('Use model {}'.format(model_name))


def inference(model_name):
    ckpt_path = os.path.join(os.getcwd(), model_name)

    cap = cv2.VideoCapture(1)
    ret, frame = cap.read()

    tf.reset_default_graph()
    image_size = max(frame.shape) // 128 * 128
    driver = inference.ServingDriver(model_name, ckpt_path, image_size=image_size)
    while True:
        start = time.time()
        ret, frame = cap.read()
        frame = np.rollaxis(frame, 0, 1)
        pred = driver.serve_images([frame])
        end = time.time()
        print("Fps: %f" % (1.0 / (end - start)))
        frame = driver.visualize(frame, pred[0])
        cv2.imshow("Image", frame)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Camera inference for efficient det networks')
    parser.add_argument('--model_type', help='Model to run: 0-6', type=int)
    args = parser.parse_args()
    model_name = 'efficientdet-d{}'.format(args.model_type)
    maybe_download(model_name)
    inference(model_name)

Hi, thank you for sharing your code, what FPS do you get with this code?
I got 0.8 FPS
best regards

Hi @jens25 , your code use GPU?

Hey @juanmanuelrq,
Unfortunately I don't have a GPU to test it, but it should use the GPU.
On my Laptop (Intel i5) with the efficientdet-d0 Network I get around 2.7 fps.

Hi @jens25, thank you for your answer,
2.7 fps with what size of image?
best regards

Sry, I forgot 640x640
If I set the image size to None in line 41, I get 3-4 fps with the d0 network.

Thanks @jens25