TokBench/eval_face.py at main · wjf5203/TokBench · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
#!/usr/bin/env python3
# Copyright (c) 2024 Junfeng Wu, Dongliang Luo. All Rights Reserved.
"""
TokBench Evaluation Script.
"""

import cv2
import numpy as np
import insightface
from insightface.app import FaceAnalysis
from insightface.data import get_image as ins_get_image
import json
import os
import argparse
from tqdm import tqdm
import imageio


def get_args_parser():
    parser = argparse.ArgumentParser('Set ', add_help=False)
    parser.add_argument('--original_image_path', type=str, default='path/to/TokBench/images/face_data/')
    parser.add_argument('--reconstruction_image_path', type=str, default='path/to/reconsturctions/face_data/chameleon/face_256/')
    parser.add_argument('--tokenizer', type=str, default='chameleon')
    parser.add_argument('--setting', type=str, choices=["256","512","1024" "480"], default='256')
    parser.add_argument("--data_type", type=str, default="image", choices=["image","video"], help=" eval for image or video")
    parser.add_argument('--meta_path', type=str, default='face_meta.json')
    parser.add_argument('--save_dir', type=str, default='output')
    return parser

def main(args):

    app = FaceAnalysis( name="antelopev2", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
    app.prepare(ctx_id=0)

    all_images = json.load(open(args.meta_path,'r'))

    eval_results = {}
    all_similarity = []
    if args.data_type == "image":
        for image_meta in tqdm(all_images):
            image_path = image_meta['image_name']
            img_ori = cv2.imread(os.path.join(args.original_image_path,image_path))
            img_rec = cv2.imread(os.path.join(args.reconstruction_image_path,image_path))
            eval_results[image_path]={
                "file_name": image_path,
                "height": image_meta['img_height'],
                "width": image_meta['img_width'],
                "results": []
                }
            for face in image_meta['faces']:
                face_input = insightface.app.common.Face(kps=np.array(face['kps']))
                embed1 = app.models['recognition'].get(img_ori, face_input)
                embed2 = app.models['recognition'].get(img_rec, face_input)
                similarity = np.dot(embed1, embed2) / (np.linalg.norm(embed1) * np.linalg.norm(embed2))
                all_similarity.append(similarity)
                eval_results[image_path]["results"].append( {"ratio":face["ratio"],
                                                            "similarity":float(similarity),
                                                            "box":face["box"],
                                                            "gt":face["gt"],
                                                            } )
    elif args.data_type == "video":
        for image_meta in tqdm(all_images):
            image_path = image_meta['video_name']
            video_reader = imageio.get_reader(os.path.join(args.original_image_path,image_path), "ffmpeg")
            ori_frames = [ frame for frame in video_reader]
            video_reader.close()

            video_reader = imageio.get_reader(os.path.join(args.reconstruction_image_path,image_path), "ffmpeg")
            rec_frames = [ frame for frame in video_reader]
            video_reader.close()

            assert len(ori_frames)==len(rec_frames)
            eval_results[image_path]={
                "file_name": image_path,
                "height": image_meta['video_height'],
                "width": image_meta['video_width'],
                "results": []
                }


            for frame_results, oriframe, recframe in zip(image_meta['frame_anns'],ori_frames,rec_frames):
                for face in frame_results:
                    face_input = insightface.app.common.Face(kps=np.array(face['kps']))
                    img_ori = oriframe[:,:,::-1]
                    img_rec = recframe[:,:,::-1]
                    embed1 = app.models['recognition'].get(img_ori, face_input)
                    embed2 = app.models['recognition'].get(img_rec, face_input)
                    distance = np.dot(embed1, embed2) / (np.linalg.norm(embed1) * np.linalg.norm(embed2))
                    all_similarity.append(distance)
                    eval_results[image_path]["results"].append( {"ratio":face["ratio"],
                                                                "similarity":float(distance),
                                                                } )
    else:
        raise NotImplementedError("undefined data type")

    mean_similarity = np.mean(all_similarity)
    mean_similarity = float(mean_similarity)


    result_log_path = os.path.join(args.save_dir, "face.json" )
    if os.path.exists(result_log_path):
        with open(result_log_path, 'r') as fp:
            log_results = json.load(fp)
    else:
        log_results = dict()

    method_name = args.tokenizer
    method_setting = args.setting
    if method_name not in log_results:
        log_results[method_name] = dict()
    log_results[method_name][method_setting] = dict(
        results=dict(
            mean_similarity = mean_similarity
        ),
        details= eval_results
    )

    with open(result_log_path, 'w') as fp:
        json.dump(log_results, fp, indent=2)


if __name__ == '__main__':
    parser = argparse.ArgumentParser('image path check script', parents=[get_args_parser()])
    args = parser.parse_args()
    main(args)