inference problem

[flynnamy]

It shows that the results is good in dataset,but it depends on scale, center_w, center_h.If I want to do inference of an image without information of scale, center_w, center_h, how do I get scale and center_w,center_h accurately.if these variable are not accurate,the landmark is not good.So I must get accurate scale, center_w and center_h. Any suggestions?

[flynnamy]

It calculates scale, center_w, center_h depends groudth of points.But if we just want to do inference an image without groundth info, this algorithm how to work?Because we can not get scale, center_w, center_h.

[lfxx]

Maybe you need a face detector to locate the target face and then feed its location into network

[flynnamy]

@lfxx It can not do like this way.You can try use a detector to get box and scale.Then you can find the points of landmark is not right.

[lfxx]

Do you mean that inference single image in wild?Use a detector to get the target face's location is useful.Here is my scrpts ,let me know if it works!@flynnamy

# ------------------------------------------------------------------------------
# Created by Gaofeng(lfxx1994@gmail.com)
# ------------------------------------------------------------------------------

import os
import argparse

import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import sys
import cv2

sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
import lib.models as models
from lib.config import config, update_config
from PIL import Image
import numpy as np
from lib.utils.transforms import crop
from lib.core.evaluation import decode_preds


def parse_args():
    parser = argparse.ArgumentParser(description='Train Face Alignment')

    parser.add_argument('--cfg',
                        default='experiments/300w/face_alignment_300w_hrnet_w18.yaml',
                        help='experiment configuration filename', type=str)
    parser.add_argument('--model-file', help='model parameters',
                        default='HR18-300W.pth', type=str)
    parser.add_argument('--imagepath', help='Path of the image to be detected', default='111.jpg',
                        type=str)
    parser.add_argument('--face', nargs='+', type=float, default=[911, 1281, 1254, 1731],
                        help='The coordinate [x1,y1,x2,y2] of a face')
    args = parser.parse_args()
    update_config(config, args)
    return args


def prepare_input(image, bbox, image_size):
    """

    :param image:The path to the image to be detected
    :param bbox:The bbox of target face
    :param image_size: refers to config file
    :return:
    """
    scale = max(bbox[2] - bbox[0], bbox[3] - bbox[1]) / 200
    center_w = (bbox[0] + bbox[2]) / 2
    center_h = (bbox[1] + bbox[3]) / 2
    center = torch.Tensor([center_w, center_h])
    scale *= 1.25
    img = np.array(Image.open(image).convert('RGB'), dtype=np.float32)
    mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
    std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
    img = crop(img, center, scale, image_size, rot=0)
    img = img.astype(np.float32)
    img = (img / 255.0 - mean) / std
    img = img.transpose([2, 0, 1])
    img = torch.Tensor(img)
    img = img.unsqueeze(0)
    return img, center, scale


def main():
    args = parse_args()
    cudnn.benchmark = config.CUDNN.BENCHMARK
    cudnn.determinstic = config.CUDNN.DETERMINISTIC
    cudnn.enabled = config.CUDNN.ENABLED

    config.defrost()
    config.MODEL.INIT_WEIGHTS = False
    config.freeze()
    model = models.get_face_alignment_net(config)
    if config.GPUS is list:
        gpus = list(config.GPUS)
    else:
        gpus = [config.GPUS]
    model = nn.DataParallel(model, device_ids=gpus).cuda()

    # load model
    state_dict = torch.load(args.model_file)
    model.load_state_dict(state_dict)
    model.eval()
    inp, center, scale = prepare_input(args.imagepath, args.face, config.MODEL.IMAGE_SIZE)
    output = model(inp)
    score_map = output.data.cpu()
    preds = decode_preds(score_map, center, scale, [64, 64])
    preds = preds.numpy()
    cv2.namedWindow('test', 0)
    img_once = cv2.imread(args.imagepath)
    for i in preds[0, :, :]:
        cv2.circle(img_once, tuple(list(int(p) for p in i.tolist())), 2, (255, 255, 0), 1)
    cv2.imshow('test', img_once)
    if cv2.waitKey(0) == 27:
        cv2.destroyAllWindows()


if __name__ == '__main__':
    main()

At the same time,modify the line 79 of evaluation.py as follows:

    if scale is list:
        for i in range(coords.size(0)):
            preds[i] = transform_preds(coords[i], center[i], scale[i], res)
    else:
        for i in range(coords.size(0)):
            preds[i] = transform_preds(coords[i], center, scale, res)

[flynnamy]

@lfxx Yes,I mean inference an image in wild.In you scripts, you use box and scale depends on points groudtruth I know it works well.But if you do not know the information of box and scale,how to get accurate box and scale?If you use a detetcor to get these,Which detector d you use?

[flynnamy]

@lfxx Any solutions to do inference without know accurate box and scale?Even using detector,it will get large difference and I have tried.

[lfxx]

@lfxx Yes,I mean inference an image in wild.In you scripts, you use box and scale depends on points groudtruth I know it works well.But if you do not know the information of box and scale,how to get accurate box and scale?If you use a detetcor to get these,Which detector d you use?

I use RetinaFace to detect face,then feed the result into net ,it works well!

[flynnamy]

@lfxx Please try wflw model,I tried this with retinaface and a wild image.It failed! Please let me you know if it works!

[lfxx]

@lfxx Please try wflw model,I tried this with retinaface and a wild image.It failed! Please let me you know if it works!

Fix this BUG,have a try~

# ------------------------------------------------------------------------------
# Created by Gaofeng(lfxx1994@gmail.com)
# ------------------------------------------------------------------------------

import os
import argparse

import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import sys
import cv2

sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
import lib.models as models
from lib.config import config, update_config
from PIL import Image
import numpy as np
from lib.utils.transforms import crop
from lib.core.evaluation import decode_preds


def parse_args():
    parser = argparse.ArgumentParser(description='Train Face Alignment')

    parser.add_argument('--cfg',
                        default='experiments/wflw/face_alignment_wflw_hrnet_w18.yaml',
                        help='experiment configuration filename', type=str)
    parser.add_argument('--model_file', help='model parameters',
                        default='HR18-WFLW.pth', type=str)
    parser.add_argument('--model_300W', help='if model is 300W,SET TRUE else FALSE',
                        default=False, type=bool)
    parser.add_argument('--imagepath', help='Path of the image to be detected', default='111.jpg',
                        type=str)
    parser.add_argument('--face', nargs='+', type=float, default=[930, 1314,1251,1727],
                        help='The coordinate [x1,y1,x2,y2] of a face')
    args = parser.parse_args()
    update_config(config, args)
    return args


def prepare_input(image, bbox, image_size):
    """

    :param image:The path of the image to be detected
    :param bbox:The bbox of target face
    :param image_size: refers to config file
    :return:
    """
    scale = max(bbox[2] - bbox[0], bbox[3] - bbox[1]) / 200
    center_w = (bbox[0] + bbox[2]) / 2
    center_h = (bbox[1] + bbox[3]) / 2
    center = torch.Tensor([center_w, center_h])
    scale *= 1.25
    img = np.array(Image.open(image).convert('RGB'), dtype=np.float32)
    mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
    std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
    img = crop(img, center, scale, image_size, rot=0)
    img = img.astype(np.float32)
    img = (img / 255.0 - mean) / std
    img = img.transpose([2, 0, 1])
    img = torch.Tensor(img)
    img = img.unsqueeze(0)
    return img, center, scale


def main():
    args = parse_args()
    cudnn.benchmark = config.CUDNN.BENCHMARK
    cudnn.determinstic = config.CUDNN.DETERMINISTIC
    cudnn.enabled = config.CUDNN.ENABLED

    config.defrost()
    config.MODEL.INIT_WEIGHTS = False
    config.freeze()
    model = models.get_face_alignment_net(config)
    if config.GPUS is list:
        gpus = list(config.GPUS)
    else:
        gpus = [config.GPUS]
    model = nn.DataParallel(model, device_ids=gpus).cuda()

    # load model
    state_dict = torch.load(args.model_file)
    if args.model_300W:
        model.load_state_dict(state_dict)
    else:
        model.module.load_state_dict(state_dict)

    model.eval()
    inp, center, scale = prepare_input(args.imagepath, args.face, config.MODEL.IMAGE_SIZE)
    output = model(inp)
    score_map = output.data.cpu()
    preds = decode_preds(score_map, center, scale, [64, 64])
    preds = preds.numpy()
    cv2.namedWindow('test', 0)
    img_once = cv2.imread(args.imagepath)
    for i in preds[0, :, :]:
        cv2.circle(img_once, tuple(list(int(p) for p in i.tolist())), 2, (255, 255, 0), 1)
    cv2.imshow('test', img_once)
    if cv2.waitKey(0) == 27:
        cv2.destroyAllWindows()


if __name__ == '__main__':
    main()

[flynnamy]

@lfxx I made a mistake of scale calculation and it works.Thanks for your reply!

[alarafat]

@lfxx Would it be possible for you to add this as a demo or evaluation script in the project? It will be helpful for anyone who wants to try the model for quick try. However, thanks for this script :)