def evaluate_img_google()

in detic/evaluation/oideval.py [0:0]

• 83 lines of code
• 28 McCabe index (conditional complexity)

    def evaluate_img_google(self, img_id, cat_id, area_rng):
        gt, dt = self._get_gt_dt(img_id, cat_id)
        if len(gt) == 0 and len(dt) == 0:
            return None
        
        if len(dt) == 0:
            return {
                "image_id": img_id,
                "category_id": cat_id,
                "area_rng": area_rng,
                "dt_ids": [],
                "dt_matches": np.array([], dtype=np.int32).reshape(1, -1),
                "dt_scores": [],
                "dt_ignore": np.array([], dtype=np.int32).reshape(1, -1),
                'num_gt': len(gt)
            }

        no_crowd_inds = [i for i, g in enumerate(gt) \
            if ('iscrowd' not in g) or g['iscrowd'] == 0]
        crowd_inds = [i for i, g in enumerate(gt) \
            if 'iscrowd' in g and g['iscrowd'] == 1]
        dt_idx = np.argsort([-d["score"] for d in dt], kind="mergesort")

        if len(self.ious[img_id, cat_id]) > 0:
            ious = self.ious[img_id, cat_id]
            iou = ious[:, no_crowd_inds]
            iou = iou[dt_idx]
            ioa = ious[:, crowd_inds]
            ioa = ioa[dt_idx]
        else:
            iou = np.zeros((len(dt_idx), 0))
            ioa = np.zeros((len(dt_idx), 0))
        scores = np.array([dt[i]['score'] for i in dt_idx])

        num_detected_boxes = len(dt)
        tp_fp_labels = np.zeros(num_detected_boxes, dtype=bool)
        is_matched_to_group_of = np.zeros(num_detected_boxes, dtype=bool)

        def compute_match_iou(iou):
            max_overlap_gt_ids = np.argmax(iou, axis=1)
            is_gt_detected = np.zeros(iou.shape[1], dtype=bool)
            for i in range(num_detected_boxes):
                gt_id = max_overlap_gt_ids[i]
                is_evaluatable = (not tp_fp_labels[i] and
                                iou[i, gt_id] >= 0.5 and
                                not is_matched_to_group_of[i])
                if is_evaluatable:
                    if not is_gt_detected[gt_id]:
                        tp_fp_labels[i] = True
                        is_gt_detected[gt_id] = True

        def compute_match_ioa(ioa):
            scores_group_of = np.zeros(ioa.shape[1], dtype=float)
            tp_fp_labels_group_of = np.ones(
                ioa.shape[1], dtype=float)
            max_overlap_group_of_gt_ids = np.argmax(ioa, axis=1)
            for i in range(num_detected_boxes):
                gt_id = max_overlap_group_of_gt_ids[i]
                is_evaluatable = (not tp_fp_labels[i] and
                                ioa[i, gt_id] >= 0.5 and
                                not is_matched_to_group_of[i])
                if is_evaluatable:
                    is_matched_to_group_of[i] = True
                    scores_group_of[gt_id] = max(scores_group_of[gt_id], scores[i])
            selector = np.where((scores_group_of > 0) & (tp_fp_labels_group_of > 0))
            scores_group_of = scores_group_of[selector]
            tp_fp_labels_group_of = tp_fp_labels_group_of[selector]

            return scores_group_of, tp_fp_labels_group_of

        if iou.shape[1] > 0:
            compute_match_iou(iou)

        scores_box_group_of = np.ndarray([0], dtype=float)
        tp_fp_labels_box_group_of = np.ndarray([0], dtype=float)

        if ioa.shape[1] > 0:
            scores_box_group_of, tp_fp_labels_box_group_of = compute_match_ioa(ioa)

        valid_entries = (~is_matched_to_group_of)

        scores =  np.concatenate(
            (scores[valid_entries], scores_box_group_of))
        tp_fps = np.concatenate(
            (tp_fp_labels[valid_entries].astype(float),
             tp_fp_labels_box_group_of))
    
        return {
            "image_id": img_id,
            "category_id": cat_id,
            "area_rng": area_rng,
            "dt_matches": np.array([1 if x > 0 else 0 for x in tp_fps], dtype=np.int32).reshape(1, -1),
            "dt_scores": [x for x in scores],
            "dt_ignore":  np.array([0 for x in scores], dtype=np.int32).reshape(1, -1),
            'num_gt': len(gt)
        }