# Copyright (c) OpenMMLab. All rights reserved. from typing import Dict, List, Optional, Sequence import numpy as np import torch from mmengine.evaluator import BaseMetric from mmengine.logging import MMLogger from shapely.geometry import LineString, Point from mmocr.registry import METRICS @METRICS.register_module() class E2EPointMetric(BaseMetric): """Point metric for textspotting. Proposed in SPTS. Args: text_score_thrs (dict): Best text score threshold searching space. Defaults to dict(start=0.8, stop=1, step=0.01). TODO: docstr collect_device (str): Device name used for collecting results from different ranks during distributed training. Must be 'cpu' or 'gpu'. Defaults to 'cpu'. prefix (str, optional): The prefix that will be added in the metric names to disambiguate homonymous metrics of different evaluators. If prefix is not provided in the argument, self.default_prefix will be used instead. Defaults to None """ default_prefix: Optional[str] = 'e2e_icdar' def __init__(self, text_score_thrs: Dict = dict(start=0.8, stop=1, step=0.01), word_spotting: bool = False, collect_device: str = 'cpu', prefix: Optional[str] = None) -> None: super().__init__(collect_device=collect_device, prefix=prefix) self.text_score_thrs = np.arange(**text_score_thrs) self.word_spotting = word_spotting def poly_center(self, poly_pts): poly_pts = np.array(poly_pts).reshape(-1, 2) num_points = poly_pts.shape[0] line1 = LineString(poly_pts[int(num_points / 2):]) line2 = LineString(poly_pts[:int(num_points / 2)]) mid_pt1 = np.array(line1.interpolate(0.5, normalized=True).coords[0]) mid_pt2 = np.array(line2.interpolate(0.5, normalized=True).coords[0]) return (mid_pt1 + mid_pt2) / 2 def process(self, data_batch: Sequence[Dict], data_samples: Sequence[Dict]) -> None: """Process one batch of data samples and predictions. The processed results should be stored in ``self.results``, which will be used to compute the metrics when all batches have been processed. Args: data_batch (Sequence[Dict]): A batch of data from dataloader. data_samples (Sequence[Dict]): A batch of outputs from the model. """ for data_sample in data_samples: pred_instances = data_sample.get('pred_instances') pred_points = pred_instances.get('points') text_scores = pred_instances.get('text_scores') if isinstance(text_scores, torch.Tensor): text_scores = text_scores.cpu().numpy() text_scores = np.array(text_scores, dtype=np.float32) pred_texts = pred_instances.get('texts') gt_instances = data_sample.get('gt_instances') gt_polys = gt_instances.get('polygons') gt_ignore_flags = gt_instances.get('ignored') gt_texts = gt_instances.get('texts') if isinstance(gt_ignore_flags, torch.Tensor): gt_ignore_flags = gt_ignore_flags.cpu().numpy() gt_points = [self.poly_center(poly) for poly in gt_polys] if self.word_spotting: gt_ignore_flags, gt_texts = self._word_spotting_filter( gt_ignore_flags, gt_texts) pred_ignore_flags = text_scores < self.text_score_thrs.min() text_scores = text_scores[~pred_ignore_flags] pred_texts = self._get_true_elements(pred_texts, ~pred_ignore_flags) pred_points = self._get_true_elements(pred_points, ~pred_ignore_flags) result = dict( text_scores=text_scores, pred_points=pred_points, gt_points=gt_points, pred_texts=pred_texts, gt_texts=gt_texts, gt_ignore_flags=gt_ignore_flags) self.results.append(result) def _get_true_elements(self, array: List, flags: np.ndarray) -> List: return [array[i] for i in self._true_indexes(flags)] def compute_metrics(self, results: List[Dict]) -> Dict: """Compute the metrics from processed results. Args: results (list[dict]): The processed results of each batch. Returns: dict: The computed metrics. The keys are the names of the metrics, and the values are corresponding results. """ logger: MMLogger = MMLogger.get_current_instance() best_eval_results = dict(hmean=-1) num_thres = len(self.text_score_thrs) num_preds = np.zeros( num_thres, dtype=int) # the number of points actually predicted num_tp = np.zeros(num_thres, dtype=int) # number of true positives num_gts = np.zeros(num_thres, dtype=int) # number of valid gts for result in results: text_scores = result['text_scores'] pred_points = result['pred_points'] gt_points = result['gt_points'] gt_texts = result['gt_texts'] pred_texts = result['pred_texts'] gt_ignore_flags = result['gt_ignore_flags'] # Filter out predictions by IoU threshold for i, text_score_thr in enumerate(self.text_score_thrs): pred_ignore_flags = text_scores < text_score_thr filtered_pred_texts = self._get_true_elements( pred_texts, ~pred_ignore_flags) filtered_pred_points = self._get_true_elements( pred_points, ~pred_ignore_flags) gt_matched = np.zeros(len(gt_texts), dtype=bool) num_gt = len(gt_texts) - np.sum(gt_ignore_flags) if num_gt == 0: continue num_gts[i] += num_gt for pred_text, pred_point in zip(filtered_pred_texts, filtered_pred_points): dists = [ Point(pred_point).distance(Point(gt_point)) for gt_point in gt_points ] min_idx = np.argmin(dists) if gt_texts[min_idx] == '###' or gt_ignore_flags[min_idx]: continue # if not gt_matched[min_idx] and self.text_match( # gt_texts[min_idx].upper(), pred_text.upper()): if (not gt_matched[min_idx] and gt_texts[min_idx].upper() == pred_text.upper()): gt_matched[min_idx] = True num_tp[i] += 1 num_preds[i] += 1 for i, text_score_thr in enumerate(self.text_score_thrs): if num_preds[i] == 0 or num_tp[i] == 0: recall, precision, hmean = 0, 0, 0 else: recall = num_tp[i] / num_gts[i] precision = num_tp[i] / num_preds[i] hmean = 2 * recall * precision / (recall + precision) eval_results = dict( precision=precision, recall=recall, hmean=hmean) logger.info(f'text score threshold: {text_score_thr:.2f}, ' f'recall: {eval_results["recall"]:.4f}, ' f'precision: {eval_results["precision"]:.4f}, ' f'hmean: {eval_results["hmean"]:.4f}\n') if eval_results['hmean'] > best_eval_results['hmean']: best_eval_results = eval_results return best_eval_results def _true_indexes(self, array: np.ndarray) -> np.ndarray: """Get indexes of True elements from a 1D boolean array.""" return np.where(array)[0]