fcos

#!/usr/bin/env python
from __future__ import print_function
from fcos import FCOS
import cv2
import skimage.io as io
import argparse
import torch
import time


def pretty_print(bbox_results):
    max_label_name_len = max([len(_["label_name"]) for _ in bbox_results])
    for item in bbox_results:
        print("{}    confidence: {:.2f}    ".format(
            item["label_name"].ljust(max_label_name_len),
            item["score"],
        ), end="")
        print("bbox: {:.1f} {:.1f} {:.1f} {:.1f}".format(
            item["box"][0],
            item["box"][1],
            item["box"][2],
            item["box"][3],
        ))


parser = argparse.ArgumentParser(description="FCOS Object Detector")
parser.add_argument(
    "input_image",
    help="path or url to an input image",
)
args = parser.parse_args()

fcos = FCOS(
    model_name="fcos_syncbn_bs32_c128_MNV2_FPN_1x",
    nms_thresh=0.6,
    cpu_only=not torch.cuda.is_available()  # if you do not have GPUs, please set cpu_only as True
)

im = io.imread(args.input_image)
assert im.shape[-1] == 3, "only 3-channel images are supported"

# convert from RGB to BGR because fcos assumes the BRG input image
im = im[..., ::-1].copy()

# resize image to have its shorter size == 800
f = 800.0 / float(min(im.shape[:2]))
im = cv2.resize(im, (0, 0), fx=f, fy=f)

start_time = time.time()

bbox_results = fcos.detect(im)

inference_time = time.time() - start_time

pretty_print(bbox_results)
print("Predicted in {:.2f} seconds.".format(inference_time))
print("Press any key to exit...")
fcos.show_bboxes(im, bbox_results)