File size: 6,427 Bytes
583456e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
# Copyright (c) Facebook, Inc. and its affiliates.
# Copyright (c) Meta Platforms, Inc. All Rights Reserved

import itertools
import json
import numpy as np
import os
from collections import OrderedDict
import PIL.Image as Image
import torch

from detectron2.data import DatasetCatalog, MetadataCatalog
from detectron2.utils.comm import all_gather, is_main_process, synchronize
from detectron2.utils.file_io import PathManager

from detectron2.evaluation import SemSegEvaluator


class GeneralizedSemSegEvaluator(SemSegEvaluator):
    """
    Evaluate semantic segmentation metrics.
    """

    def __init__(
        self,
        dataset_name,
        distributed=True,
        output_dir=None,
        *,
        num_classes=None,
        ignore_label=None,
        post_process_func=None,
    ):
        super().__init__(
            dataset_name,
            distributed=distributed,
            output_dir=output_dir,
            num_classes=num_classes,
            ignore_label=ignore_label,
        )
        meta = MetadataCatalog.get(dataset_name)
        try:
            self._evaluation_set = meta.evaluation_set
        except AttributeError:
            self._evaluation_set = None
        self.post_process_func = (
            post_process_func
            if post_process_func is not None
            else lambda x, **kwargs: x
        )

    def process(self, inputs, outputs):
        """
        Args:
            inputs: the inputs to a model.
                It is a list of dicts. Each dict corresponds to an image and
                contains keys like "height", "width", "file_name".
            outputs: the outputs of a model. It is either list of semantic segmentation predictions
                (Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic
                segmentation prediction in the same format.
        """
        for input, output in zip(inputs, outputs):
            output = self.post_process_func(
                output["sem_seg"], image=np.array(Image.open(input["file_name"]))
            )
            output = output.argmax(dim=0).to(self._cpu_device)
            pred = np.array(output, dtype=np.int)
            with PathManager.open(
                self.input_file_to_gt_file[input["file_name"]], "rb"
            ) as f:
                gt = np.array(Image.open(f), dtype=np.int)

            gt[gt == self._ignore_label] = self._num_classes

            self._conf_matrix += np.bincount(
                (self._num_classes + 1) * pred.reshape(-1) + gt.reshape(-1),
                minlength=self._conf_matrix.size,
            ).reshape(self._conf_matrix.shape)

            self._predictions.extend(self.encode_json_sem_seg(pred, input["file_name"]))

    def evaluate(self):
        """
        Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval):

        * Mean intersection-over-union averaged across classes (mIoU)
        * Frequency Weighted IoU (fwIoU)
        * Mean pixel accuracy averaged across classes (mACC)
        * Pixel Accuracy (pACC)
        """
        if self._distributed:
            synchronize()
            conf_matrix_list = all_gather(self._conf_matrix)
            self._predictions = all_gather(self._predictions)
            self._predictions = list(itertools.chain(*self._predictions))
            if not is_main_process():
                return

            self._conf_matrix = np.zeros_like(self._conf_matrix)
            for conf_matrix in conf_matrix_list:
                self._conf_matrix += conf_matrix

        if self._output_dir:
            PathManager.mkdirs(self._output_dir)
            file_path = os.path.join(self._output_dir, "sem_seg_predictions.json")
            with PathManager.open(file_path, "w") as f:
                f.write(json.dumps(self._predictions))

        acc = np.full(self._num_classes, np.nan, dtype=np.float)
        iou = np.full(self._num_classes, np.nan, dtype=np.float)
        tp = self._conf_matrix.diagonal()[:-1].astype(np.float)
        pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float)
        class_weights = pos_gt / np.sum(pos_gt)
        pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float)
        acc_valid = pos_gt > 0
        acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid]
        iou_valid = (pos_gt + pos_pred) > 0
        union = pos_gt + pos_pred - tp
        iou[acc_valid] = tp[acc_valid] / union[acc_valid]
        macc = np.sum(acc[acc_valid]) / np.sum(acc_valid)
        miou = np.sum(iou[acc_valid]) / np.sum(iou_valid)
        fiou = np.sum(iou[acc_valid] * class_weights[acc_valid])
        pacc = np.sum(tp) / np.sum(pos_gt)

        res = {}
        res["mIoU"] = 100 * miou
        res["fwIoU"] = 100 * fiou
        for i, name in enumerate(self._class_names):
            res["IoU-{}".format(name)] = 100 * iou[i]
        res["mACC"] = 100 * macc
        res["pACC"] = 100 * pacc
        for i, name in enumerate(self._class_names):
            res["ACC-{}".format(name)] = 100 * acc[i]
        if self._evaluation_set is not None:
            for set_name, set_inds in self._evaluation_set.items():
                iou_list = []
                set_inds = np.array(set_inds, np.int)
                mask = np.zeros((len(iou),)).astype(np.bool)
                mask[set_inds] = 1
                miou = np.sum(iou[mask][acc_valid[mask]]) / np.sum(iou_valid[mask])
                pacc = np.sum(tp[mask]) / np.sum(pos_gt[mask])
                res["mIoU-{}".format(set_name)] = 100 * miou
                res["pAcc-{}".format(set_name)] = 100 * pacc
                iou_list.append(miou)
                miou = np.sum(iou[~mask][acc_valid[~mask]]) / np.sum(iou_valid[~mask])
                pacc = np.sum(tp[~mask]) / np.sum(pos_gt[~mask])
                res["mIoU-un{}".format(set_name)] = 100 * miou
                res["pAcc-un{}".format(set_name)] = 100 * pacc
                iou_list.append(miou)
                res["hIoU-{}".format(set_name)] = (
                    100 * len(iou_list) / sum([1 / iou for iou in iou_list])
                )
        if self._output_dir:
            file_path = os.path.join(self._output_dir, "sem_seg_evaluation.pth")
            with PathManager.open(file_path, "wb") as f:
                torch.save(res, f)
        results = OrderedDict({"sem_seg": res})
        self._logger.info(results)
        return results