File size: 15,593 Bytes
34d1f8b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
# Copyright (c) OpenMMLab. All rights reserved.
# adapted from https://github.com/ScanNet/ScanNet/blob/master/BenchmarkScripts/3d_evaluation/evaluate_semantic_instance.py # noqa
from copy import deepcopy

import numpy as np

from . import util_3d


def evaluate_matches(matches, class_labels, options):
    """Evaluate instance segmentation from matched gt and predicted instances
    for all scenes.

    Args:
        matches (dict): Contains gt2pred and pred2gt infos for every scene.
        class_labels (tuple[str]): Class names.
        options (dict): ScanNet evaluator options. See get_options.

    Returns:
        np.array: Average precision scores for all thresholds and categories.
    """
    overlaps = options['overlaps']
    min_region_sizes = [options['min_region_sizes'][0]]
    dist_threshes = [options['distance_threshes'][0]]
    dist_confs = [options['distance_confs'][0]]

    # results: class x overlap
    ap = np.zeros((len(dist_threshes), len(class_labels), len(overlaps)))
    for di, (min_region_size, distance_thresh, distance_conf) in enumerate(
            zip(min_region_sizes, dist_threshes, dist_confs)):
        for oi, overlap_th in enumerate(overlaps):
            pred_visited = {}
            for m in matches:
                for label_name in class_labels:
                    for p in matches[m]['pred'][label_name]:
                        if 'filename' in p:
                            pred_visited[p['filename']] = False
            for li, label_name in enumerate(class_labels):
                y_true = np.empty(0)
                y_score = np.empty(0)
                hard_false_negatives = 0
                has_gt = False
                has_pred = False
                for m in matches:
                    pred_instances = matches[m]['pred'][label_name]
                    gt_instances = matches[m]['gt'][label_name]
                    # filter groups in ground truth
                    gt_instances = [
                        gt for gt in gt_instances
                        if gt['instance_id'] >= 1000 and gt['vert_count'] >=
                        min_region_size and gt['med_dist'] <= distance_thresh
                        and gt['dist_conf'] >= distance_conf
                    ]
                    if gt_instances:
                        has_gt = True
                    if pred_instances:
                        has_pred = True

                    cur_true = np.ones(len(gt_instances))
                    cur_score = np.ones(len(gt_instances)) * (-float('inf'))
                    cur_match = np.zeros(len(gt_instances), dtype=bool)
                    # collect matches
                    for (gti, gt) in enumerate(gt_instances):
                        found_match = False
                        for pred in gt['matched_pred']:
                            # greedy assignments
                            if pred_visited[pred['filename']]:
                                continue
                            overlap = float(pred['intersection']) / (
                                gt['vert_count'] + pred['vert_count'] -
                                pred['intersection'])
                            if overlap > overlap_th:
                                confidence = pred['confidence']
                                # if already have a prediction for this gt,
                                # the prediction with the lower score is automatically a false positive # noqa
                                if cur_match[gti]:
                                    max_score = max(cur_score[gti], confidence)
                                    min_score = min(cur_score[gti], confidence)
                                    cur_score[gti] = max_score
                                    # append false positive
                                    cur_true = np.append(cur_true, 0)
                                    cur_score = np.append(cur_score, min_score)
                                    cur_match = np.append(cur_match, True)
                                # otherwise set score
                                else:
                                    found_match = True
                                    cur_match[gti] = True
                                    cur_score[gti] = confidence
                                    pred_visited[pred['filename']] = True
                        if not found_match:
                            hard_false_negatives += 1
                    # remove non-matched ground truth instances
                    cur_true = cur_true[cur_match]
                    cur_score = cur_score[cur_match]

                    # collect non-matched predictions as false positive
                    for pred in pred_instances:
                        found_gt = False
                        for gt in pred['matched_gt']:
                            overlap = float(gt['intersection']) / (
                                gt['vert_count'] + pred['vert_count'] -
                                gt['intersection'])
                            if overlap > overlap_th:
                                found_gt = True
                                break
                        if not found_gt:
                            num_ignore = pred['void_intersection']
                            for gt in pred['matched_gt']:
                                # group?
                                if gt['instance_id'] < 1000:
                                    num_ignore += gt['intersection']
                                # small ground truth instances
                                if gt['vert_count'] < min_region_size or gt[
                                        'med_dist'] > distance_thresh or gt[
                                            'dist_conf'] < distance_conf:
                                    num_ignore += gt['intersection']
                            proportion_ignore = float(
                                num_ignore) / pred['vert_count']
                            # if not ignored append false positive
                            if proportion_ignore <= overlap_th:
                                cur_true = np.append(cur_true, 0)
                                confidence = pred['confidence']
                                cur_score = np.append(cur_score, confidence)

                    # append to overall results
                    y_true = np.append(y_true, cur_true)
                    y_score = np.append(y_score, cur_score)

                # compute average precision
                if has_gt and has_pred:
                    # compute precision recall curve first

                    # sorting and cumsum
                    score_arg_sort = np.argsort(y_score)
                    y_score_sorted = y_score[score_arg_sort]
                    y_true_sorted = y_true[score_arg_sort]
                    y_true_sorted_cumsum = np.cumsum(y_true_sorted)

                    # unique thresholds
                    (thresholds, unique_indices) = np.unique(
                        y_score_sorted, return_index=True)
                    num_prec_recall = len(unique_indices) + 1

                    # prepare precision recall
                    num_examples = len(y_score_sorted)
                    # follow https://github.com/ScanNet/ScanNet/pull/26 ? # noqa
                    num_true_examples = y_true_sorted_cumsum[-1] if len(
                        y_true_sorted_cumsum) > 0 else 0
                    precision = np.zeros(num_prec_recall)
                    recall = np.zeros(num_prec_recall)

                    # deal with the first point
                    y_true_sorted_cumsum = np.append(y_true_sorted_cumsum, 0)
                    # deal with remaining
                    for idx_res, idx_scores in enumerate(unique_indices):
                        cumsum = y_true_sorted_cumsum[idx_scores - 1]
                        tp = num_true_examples - cumsum
                        fp = num_examples - idx_scores - tp
                        fn = cumsum + hard_false_negatives
                        p = float(tp) / (tp + fp)
                        r = float(tp) / (tp + fn)
                        precision[idx_res] = p
                        recall[idx_res] = r

                    # first point in curve is artificial
                    precision[-1] = 1.
                    recall[-1] = 0.

                    # compute average of precision-recall curve
                    recall_for_conv = np.copy(recall)
                    recall_for_conv = np.append(recall_for_conv[0],
                                                recall_for_conv)
                    recall_for_conv = np.append(recall_for_conv, 0.)

                    stepWidths = np.convolve(recall_for_conv, [-0.5, 0, 0.5],
                                             'valid')
                    # integrate is now simply a dot product
                    ap_current = np.dot(precision, stepWidths)

                elif has_gt:
                    ap_current = 0.0
                else:
                    ap_current = float('nan')
                ap[di, li, oi] = ap_current
    return ap


def compute_averages(aps, options, class_labels):
    """Averages AP scores for all categories.

    Args:
        aps (np.array): AP scores for all thresholds and categories.
        options (dict): ScanNet evaluator options. See get_options.
        class_labels (tuple[str]): Class names.

    Returns:
        dict: Overall and per-category AP scores.
    """
    d_inf = 0
    o50 = np.where(np.isclose(options['overlaps'], 0.5))
    o25 = np.where(np.isclose(options['overlaps'], 0.25))
    o_all_but25 = np.where(
        np.logical_not(np.isclose(options['overlaps'], 0.25)))
    avg_dict = {}
    avg_dict['all_ap'] = np.nanmean(aps[d_inf, :, o_all_but25])
    avg_dict['all_ap_50%'] = np.nanmean(aps[d_inf, :, o50])
    avg_dict['all_ap_25%'] = np.nanmean(aps[d_inf, :, o25])
    avg_dict['classes'] = {}
    for (li, label_name) in enumerate(class_labels):
        avg_dict['classes'][label_name] = {}
        avg_dict['classes'][label_name]['ap'] = np.average(aps[d_inf, li,
                                                               o_all_but25])
        avg_dict['classes'][label_name]['ap50%'] = np.average(aps[d_inf, li,
                                                                  o50])
        avg_dict['classes'][label_name]['ap25%'] = np.average(aps[d_inf, li,
                                                                  o25])
    return avg_dict


def assign_instances_for_scan(pred_info, gt_ids, options, valid_class_ids,
                              class_labels, id_to_label):
    """Assign gt and predicted instances for a single scene.

    Args:
        pred_info (dict): Predicted masks, labels and scores.
        gt_ids (np.array): Ground truth instance masks.
        options (dict): ScanNet evaluator options. See get_options.
        valid_class_ids (tuple[int]): Ids of valid categories.
        class_labels (tuple[str]): Class names.
        id_to_label (dict[int, str]): Mapping of valid class id to class label.

    Returns:
        dict: Per class assigned gt to predicted instances.
        dict: Per class assigned predicted to gt instances.
    """
    # get gt instances
    gt_instances = util_3d.get_instances(gt_ids, valid_class_ids, class_labels,
                                         id_to_label)
    # associate
    gt2pred = deepcopy(gt_instances)
    for label in gt2pred:
        for gt in gt2pred[label]:
            gt['matched_pred'] = []
    pred2gt = {}
    for label in class_labels:
        pred2gt[label] = []
    num_pred_instances = 0
    # mask of void labels in the ground truth
    bool_void = np.logical_not(np.in1d(gt_ids // 1000, valid_class_ids))
    # go through all prediction masks
    for pred_mask_file in pred_info:
        label_id = int(pred_info[pred_mask_file]['label_id'])
        conf = pred_info[pred_mask_file]['conf']
        if not label_id in id_to_label:  # noqa E713
            continue
        label_name = id_to_label[label_id]
        # read the mask
        pred_mask = pred_info[pred_mask_file]['mask']
        if len(pred_mask) != len(gt_ids):
            raise ValueError('len(pred_mask) != len(gt_ids)')
        # convert to binary
        pred_mask = np.not_equal(pred_mask, 0)
        num = np.count_nonzero(pred_mask)
        if num < options['min_region_sizes'][0]:
            continue  # skip if empty

        pred_instance = {}
        pred_instance['filename'] = pred_mask_file
        pred_instance['pred_id'] = num_pred_instances
        pred_instance['label_id'] = label_id
        pred_instance['vert_count'] = num
        pred_instance['confidence'] = conf
        pred_instance['void_intersection'] = np.count_nonzero(
            np.logical_and(bool_void, pred_mask))

        # matched gt instances
        matched_gt = []
        # go through all gt instances with matching label
        for (gt_num, gt_inst) in enumerate(gt2pred[label_name]):
            intersection = np.count_nonzero(
                np.logical_and(gt_ids == gt_inst['instance_id'], pred_mask))
            if intersection > 0:
                gt_copy = gt_inst.copy()
                pred_copy = pred_instance.copy()
                gt_copy['intersection'] = intersection
                pred_copy['intersection'] = intersection
                matched_gt.append(gt_copy)
                gt2pred[label_name][gt_num]['matched_pred'].append(pred_copy)
        pred_instance['matched_gt'] = matched_gt
        num_pred_instances += 1
        pred2gt[label_name].append(pred_instance)

    return gt2pred, pred2gt


def scannet_eval(preds, gts, options, valid_class_ids, class_labels,
                 id_to_label):
    """Evaluate instance segmentation in ScanNet protocol.

    Args:
        preds (list[dict]): Per scene predictions of mask, label and
            confidence.
        gts (list[np.array]): Per scene ground truth instance masks.
        options (dict): ScanNet evaluator options. See get_options.
        valid_class_ids (tuple[int]): Ids of valid categories.
        class_labels (tuple[str]): Class names.
        id_to_label (dict[int, str]): Mapping of valid class id to class label.

    Returns:
        dict: Overall and per-category AP scores.
    """
    options = get_options(options)
    matches = {}
    for i, (pred, gt) in enumerate(zip(preds, gts)):
        matches_key = i
        # assign gt to predictions
        gt2pred, pred2gt = assign_instances_for_scan(pred, gt, options,
                                                     valid_class_ids,
                                                     class_labels, id_to_label)
        matches[matches_key] = {}
        matches[matches_key]['gt'] = gt2pred
        matches[matches_key]['pred'] = pred2gt

    ap_scores = evaluate_matches(matches, class_labels, options)
    avgs = compute_averages(ap_scores, options, class_labels)
    return avgs


def get_options(options=None):
    """Set ScanNet evaluator options.

    Args:
        options (dict, optional): Not default options. Default: None.

    Returns:
        dict: Updated options with all 4 keys.
    """
    assert options is None or isinstance(options, dict)
    _options = dict(
        overlaps=np.append(np.arange(0.5, 0.95, 0.05), 0.25),
        min_region_sizes=np.array([100]),
        distance_threshes=np.array([float('inf')]),
        distance_confs=np.array([-float('inf')]))
    if options is not None:
        _options.update(options)
    return _options