valor-lite 0.33.3__py3-none-any.whl → 0.33.5__py3-none-any.whl

valor_lite/detection/manager.py

@@ -3,12 +3,20 @@ from dataclasses import dataclass
 
 import numpy as np
 from numpy.typing import NDArray
+from shapely.geometry import Polygon as ShapelyPolygon
 from tqdm import tqdm
-from valor_lite.detection.annotation import Detection
+from valor_lite.detection.annotation import (
+    Bitmask,
+    BoundingBox,
+    Detection,
+    Polygon,
+)
 from valor_lite.detection.computation import (
-    compute_detailed_counts,
-    compute_iou,
+    compute_bbox_iou,
+    compute_bitmask_iou,
+    compute_confusion_matrix,
     compute_metrics,
+    compute_polygon_iou,
     compute_ranked_pairs,
 )
 from valor_lite.detection.metric import (
@@ -18,8 +26,8 @@ from valor_lite.detection.metric import (
     Accuracy,
     APAveragedOverIOUs,
     ARAveragedOverScores,
+    ConfusionMatrix,
     Counts,
-    DetailedCounts,
     MetricType,
     Precision,
     PrecisionRecallCurve,
@@ -35,7 +43,7 @@ Usage
 -----
 
 loader = DataLoader()
-loader.add_data(
+loader.add_bounding_boxes(
     groundtruths=groundtruths,
     predictions=predictions,
 )
@@ -51,9 +59,107 @@ filtered_metrics = evaluator.evaluate(iou_thresholds=[0.5], filter_mask=filter_m
 """
 
 
+def _get_valor_dict_annotation_key(
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask],
+) -> str:
+    """Get the correct JSON key to extract a given annotation type."""
+
+    if issubclass(annotation_type, BoundingBox):
+        return "bounding_box"
+    if issubclass(annotation_type, Polygon):
+        return "polygon"
+    else:
+        return "raster"
+
+
+def _get_annotation_representation(
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask],
+) -> str:
+    """Get the correct representation of an annotation object."""
+
+    representation = (
+        "extrema"
+        if issubclass(annotation_type, BoundingBox)
+        else ("mask" if issubclass(annotation_type, Bitmask) else "shape")
+    )
+
+    return representation
+
+
+def _get_annotation_representation_from_valor_dict(
+    data: list,
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask],
+) -> tuple[float, float, float, float] | ShapelyPolygon | NDArray[np.bool_]:
+    """Get the correct representation of an annotation object from a valor dictionary."""
+
+    if issubclass(annotation_type, BoundingBox):
+        x = [point[0] for shape in data for point in shape]
+        y = [point[1] for shape in data for point in shape]
+        return (min(x), max(x), min(y), max(y))
+    if issubclass(annotation_type, Polygon):
+        return ShapelyPolygon(data)
+    else:
+        return np.array(data)
+
+
+def _get_annotation_data(
+    keyed_groundtruths: dict,
+    keyed_predictions: dict,
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask] | None,
+    key=int,
+) -> np.ndarray:
+    """Create an array of annotation pairs for use when calculating IOU. Needed because we unpack bounding box representations, but not bitmask or polygon representations."""
+    if annotation_type == BoundingBox:
+        return np.array(
+            [
+                np.array([*gextrema, *pextrema])
+                for _, _, _, pextrema in keyed_predictions[key]
+                for _, _, gextrema in keyed_groundtruths[key]
+            ]
+        )
+    else:
+        return np.array(
+            [
+                np.array([groundtruth_obj, prediction_obj])
+                for _, _, _, prediction_obj in keyed_predictions[key]
+                for _, _, groundtruth_obj in keyed_groundtruths[key]
+            ]
+        )
+
+
+def compute_iou(
+    data: NDArray[np.floating],
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask],
+) -> NDArray[np.floating]:
+    """
+    Computes intersection-over-union (IoU) calculations for various annotation types.
+
+    Parameters
+    ----------
+    data : NDArray[np.floating]
+        A sorted array of bounding box, bitmask, or polygon pairs.
+    annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask]
+        The type of annotation contained in the data.
+
+
+    Returns
+    -------
+    NDArray[np.floating]
+        Computed IoU's.
+    """
+
+    if annotation_type == BoundingBox:
+        return compute_bbox_iou(data=data)
+    elif annotation_type == Bitmask:
+        return compute_bitmask_iou(data=data)
+    else:
+        return compute_polygon_iou(data=data)
+
+
 @dataclass
 class Filter:
-    indices: NDArray[np.int32]
+    ranked_indices: NDArray[np.int32]
+    detailed_indices: NDArray[np.int32]
     label_metadata: NDArray[np.int32]
 
 
@@ -152,12 +258,14 @@ class Evaluator:
         Filter
             A filter object that can be passed to the `evaluate` method.
         """
-        n_rows = self._ranked_pairs.shape[0]
 
         n_datums = self._label_metadata_per_datum.shape[1]
         n_labels = self._label_metadata_per_datum.shape[2]
 
-        mask_pairs = np.ones((n_rows, 1), dtype=np.bool_)
+        mask_ranked = np.ones((self._ranked_pairs.shape[0], 1), dtype=np.bool_)
+        mask_detailed = np.ones(
+            (self._detailed_pairs.shape[0], 1), dtype=np.bool_
+        )
         mask_datums = np.ones(n_datums, dtype=np.bool_)
         mask_labels = np.ones(n_labels, dtype=np.bool_)
 
@@ -167,50 +275,50 @@ class Evaluator:
                     [self.uid_to_index[uid] for uid in datum_uids],
                     dtype=np.int32,
                 )
-            mask = np.zeros_like(mask_pairs, dtype=np.bool_)
-            mask[
-                np.isin(self._ranked_pairs[:, 0].astype(int), datum_uids)
-            ] = True
-            mask_pairs &= mask
-
-            mask = np.zeros_like(mask_datums, dtype=np.bool_)
-            mask[datum_uids] = True
-            mask_datums &= mask
+            mask_ranked[
+                ~np.isin(self._ranked_pairs[:, 0].astype(int), datum_uids)
+            ] = False
+            mask_detailed[
+                ~np.isin(self._detailed_pairs[:, 0].astype(int), datum_uids)
+            ] = False
+            mask_datums[~np.isin(np.arange(n_datums), datum_uids)] = False
 
         if labels is not None:
             if isinstance(labels, list):
                 labels = np.array(
                     [self.label_to_index[label] for label in labels]
                 )
-            mask = np.zeros_like(mask_pairs, dtype=np.bool_)
-            mask[np.isin(self._ranked_pairs[:, 4].astype(int), labels)] = True
-            mask_pairs &= mask
-
-            mask = np.zeros_like(mask_labels, dtype=np.bool_)
-            mask[labels] = True
-            mask_labels &= mask
+            mask_ranked[
+                ~np.isin(self._ranked_pairs[:, 4].astype(int), labels)
+            ] = False
+            mask_detailed[
+                ~np.isin(self._detailed_pairs[:, 4].astype(int), labels)
+            ] = False
+            mask_labels[~np.isin(np.arange(n_labels), labels)] = False
 
         if label_keys is not None:
             if isinstance(label_keys, list):
                 label_keys = np.array(
                     [self.label_key_to_index[key] for key in label_keys]
                 )
-            label_indices = np.where(
-                np.isclose(self._label_metadata[:, 2], label_keys)
-            )[0]
-            mask = np.zeros_like(mask_pairs, dtype=np.bool_)
-            mask[
-                np.isin(self._ranked_pairs[:, 4].astype(int), label_indices)
-            ] = True
-            mask_pairs &= mask
-
-            mask = np.zeros_like(mask_labels, dtype=np.bool_)
-            mask[label_indices] = True
-            mask_labels &= mask
-
-        mask = mask_datums[:, np.newaxis] & mask_labels[np.newaxis, :]
+            label_indices = (
+                np.where(np.isclose(self._label_metadata[:, 2], label_keys))[0]
+                if label_keys.size > 0
+                else np.array([])
+            )
+            mask_ranked[
+                ~np.isin(self._ranked_pairs[:, 4].astype(int), label_indices)
+            ] = False
+            mask_detailed[
+                ~np.isin(self._detailed_pairs[:, 4].astype(int), label_indices)
+            ] = False
+            mask_labels[~np.isin(np.arange(n_labels), label_indices)] = False
+
+        mask_label_metadata = (
+            mask_datums[:, np.newaxis] & mask_labels[np.newaxis, :]
+        )
         label_metadata_per_datum = self._label_metadata_per_datum.copy()
-        label_metadata_per_datum[:, ~mask] = 0
+        label_metadata_per_datum[:, ~mask_label_metadata] = 0
 
         label_metadata = np.zeros_like(self._label_metadata, dtype=np.int32)
         label_metadata[:, :2] = np.transpose(
@@ -222,7 +330,8 @@ class Evaluator:
         label_metadata[:, 2] = self._label_metadata[:, 2]
 
         return Filter(
-            indices=np.where(mask_pairs)[0],
+            ranked_indices=np.where(mask_ranked)[0],
+            detailed_indices=np.where(mask_detailed)[0],
             label_metadata=label_metadata,
         )
 
@@ -246,7 +355,7 @@ class Evaluator:
         score_thresholds : list[float]
             A list of score thresholds to compute metrics over.
         number_of_examples : int, default=0
-            Number of annotation examples to return in DetailedCounts.
+            Maximum number of annotation examples to return in ConfusionMatrix.
         filter_ : Filter, optional
             An optional filter object.
 
@@ -256,10 +365,12 @@ class Evaluator:
             A dictionary mapping MetricType enumerations to lists of computed metrics.
         """
 
-        data = self._ranked_pairs
+        ranked_pairs = self._ranked_pairs
+        detailed_pairs = self._detailed_pairs
         label_metadata = self._label_metadata
         if filter_ is not None:
-            data = data[filter_.indices]
+            ranked_pairs = ranked_pairs[filter_.ranked_indices]
+            detailed_pairs = detailed_pairs[filter_.detailed_indices]
             label_metadata = filter_.label_metadata
 
         (
@@ -278,7 +389,7 @@ class Evaluator:
             precision_recall,
             pr_curves,
         ) = compute_metrics(
-            data=data,
+            data=ranked_pairs,
             label_metadata=label_metadata,
             iou_thresholds=np.array(iou_thresholds),
             score_thresholds=np.array(score_thresholds),
@@ -294,7 +405,7 @@ class Evaluator:
             )
             for iou_idx in range(average_precision.shape[0])
             for label_idx in range(average_precision.shape[1])
-            if int(label_metadata[label_idx][0]) > 0
+            if int(label_metadata[label_idx, 0]) > 0
         ]
 
         metrics[MetricType.mAP] = [
@@ -314,7 +425,7 @@ class Evaluator:
                 label=self.index_to_label[label_idx],
             )
             for label_idx in range(self.n_labels)
-            if int(label_metadata[label_idx][0]) > 0
+            if int(label_metadata[label_idx, 0]) > 0
         ]
 
         metrics[MetricType.mAPAveragedOverIOUs] = [
@@ -337,7 +448,7 @@ class Evaluator:
             )
             for score_idx in range(average_recall.shape[0])
             for label_idx in range(average_recall.shape[1])
-            if int(label_metadata[label_idx][0]) > 0
+            if int(label_metadata[label_idx, 0]) > 0
         ]
 
         metrics[MetricType.mAR] = [
@@ -359,7 +470,7 @@ class Evaluator:
                 label=self.index_to_label[label_idx],
             )
             for label_idx in range(self.n_labels)
-            if int(label_metadata[label_idx][0]) > 0
+            if int(label_metadata[label_idx, 0]) > 0
         ]
 
         metrics[MetricType.mARAveragedOverScores] = [
@@ -382,16 +493,17 @@ class Evaluator:
             )
             for iou_idx, iou_threshold in enumerate(iou_thresholds)
             for label_idx, label in self.index_to_label.items()
-            if int(label_metadata[label_idx][0]) > 0
+            if int(label_metadata[label_idx, 0]) > 0
         ]
 
         for label_idx, label in self.index_to_label.items():
+
+            if label_metadata[label_idx, 0] == 0:
+                continue
+
             for score_idx, score_threshold in enumerate(score_thresholds):
                 for iou_idx, iou_threshold in enumerate(iou_thresholds):
 
-                    if label_metadata[label_idx, 0] == 0:
-                        continue
-
                     row = precision_recall[iou_idx][score_idx][label_idx]
                     kwargs = {
                         "label": label,
@@ -432,11 +544,15 @@ class Evaluator:
                         )
                     )
 
-        if MetricType.DetailedCounts in metrics_to_return:
-            metrics[MetricType.DetailedCounts] = self._compute_detailed_counts(
+        if MetricType.ConfusionMatrix in metrics_to_return:
+            metrics[
+                MetricType.ConfusionMatrix
+            ] = self._compute_confusion_matrix(
+                data=detailed_pairs,
+                label_metadata=label_metadata,
                 iou_thresholds=iou_thresholds,
                 score_thresholds=score_thresholds,
-                n_samples=number_of_examples,
+                number_of_examples=number_of_examples,
             )
 
         for metric in set(metrics.keys()):
@@ -445,14 +561,281 @@ class Evaluator:
 
         return metrics
 
-    def _compute_detailed_counts(
+    def _unpack_confusion_matrix(
+        self,
+        confusion_matrix: NDArray[np.floating],
+        label_key_idx: int,
+        number_of_labels: int,
+        number_of_examples: int,
+    ) -> dict[
+        str,
+        dict[
+            str,
+            dict[
+                str,
+                int
+                | list[
+                    dict[
+                        str,
+                        str | float | tuple[float, float, float, float],
+                    ]
+                ],
+            ],
+        ],
+    ]:
+        """
+        Unpacks a numpy array of confusion matrix counts and examples.
+        """
+
+        datum_idx = lambda gt_label_idx, pd_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+            confusion_matrix[
+                gt_label_idx,
+                pd_label_idx,
+                example_idx * 4 + 1,
+            ]
+        )
+
+        groundtruth_idx = lambda gt_label_idx, pd_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+            confusion_matrix[
+                gt_label_idx,
+                pd_label_idx,
+                example_idx * 4 + 2,
+            ]
+        )
+
+        prediction_idx = lambda gt_label_idx, pd_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+            confusion_matrix[
+                gt_label_idx,
+                pd_label_idx,
+                example_idx * 4 + 3,
+            ]
+        )
+
+        score_idx = lambda gt_label_idx, pd_label_idx, example_idx: float(  # noqa: E731 - lambda fn
+            confusion_matrix[
+                gt_label_idx,
+                pd_label_idx,
+                example_idx * 4 + 4,
+            ]
+        )
+
+        return {
+            self.index_to_label[gt_label_idx][1]: {
+                self.index_to_label[pd_label_idx][1]: {
+                    "count": max(
+                        int(confusion_matrix[gt_label_idx, pd_label_idx, 0]),
+                        0,
+                    ),
+                    "examples": [
+                        {
+                            "datum": self.index_to_uid[
+                                datum_idx(
+                                    gt_label_idx, pd_label_idx, example_idx
+                                )
+                            ],
+                            "groundtruth": tuple(
+                                self.groundtruth_examples[
+                                    datum_idx(
+                                        gt_label_idx,
+                                        pd_label_idx,
+                                        example_idx,
+                                    )
+                                ][
+                                    groundtruth_idx(
+                                        gt_label_idx,
+                                        pd_label_idx,
+                                        example_idx,
+                                    )
+                                ].tolist()
+                            ),
+                            "prediction": tuple(
+                                self.prediction_examples[
+                                    datum_idx(
+                                        gt_label_idx,
+                                        pd_label_idx,
+                                        example_idx,
+                                    )
+                                ][
+                                    prediction_idx(
+                                        gt_label_idx,
+                                        pd_label_idx,
+                                        example_idx,
+                                    )
+                                ].tolist()
+                            ),
+                            "score": score_idx(
+                                gt_label_idx, pd_label_idx, example_idx
+                            ),
+                        }
+                        for example_idx in range(number_of_examples)
+                        if datum_idx(gt_label_idx, pd_label_idx, example_idx)
+                        >= 0
+                    ],
+                }
+                for pd_label_idx in range(number_of_labels)
+                if (
+                    self.label_index_to_label_key_index[pd_label_idx]
+                    == label_key_idx
+                )
+            }
+            for gt_label_idx in range(number_of_labels)
+            if (
+                self.label_index_to_label_key_index[gt_label_idx]
+                == label_key_idx
+            )
+        }
+
+    def _unpack_hallucinations(
+        self,
+        hallucinations: NDArray[np.floating],
+        label_key_idx: int,
+        number_of_labels: int,
+        number_of_examples: int,
+    ) -> dict[
+        str,
+        dict[
+            str,
+            int
+            | list[dict[str, str | float | tuple[float, float, float, float]]],
+        ],
+    ]:
+        """
+        Unpacks a numpy array of hallucination counts and examples.
+        """
+
+        datum_idx = (
+            lambda pd_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+                hallucinations[
+                    pd_label_idx,
+                    example_idx * 3 + 1,
+                ]
+            )
+        )
+
+        prediction_idx = (
+            lambda pd_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+                hallucinations[
+                    pd_label_idx,
+                    example_idx * 3 + 2,
+                ]
+            )
+        )
+
+        score_idx = (
+            lambda pd_label_idx, example_idx: float(  # noqa: E731 - lambda fn
+                hallucinations[
+                    pd_label_idx,
+                    example_idx * 3 + 3,
+                ]
+            )
+        )
+
+        return {
+            self.index_to_label[pd_label_idx][1]: {
+                "count": max(
+                    int(hallucinations[pd_label_idx, 0]),
+                    0,
+                ),
+                "examples": [
+                    {
+                        "datum": self.index_to_uid[
+                            datum_idx(pd_label_idx, example_idx)
+                        ],
+                        "prediction": tuple(
+                            self.prediction_examples[
+                                datum_idx(pd_label_idx, example_idx)
+                            ][
+                                prediction_idx(pd_label_idx, example_idx)
+                            ].tolist()
+                        ),
+                        "score": score_idx(pd_label_idx, example_idx),
+                    }
+                    for example_idx in range(number_of_examples)
+                    if datum_idx(pd_label_idx, example_idx) >= 0
+                ],
+            }
+            for pd_label_idx in range(number_of_labels)
+            if (
+                self.label_index_to_label_key_index[pd_label_idx]
+                == label_key_idx
+            )
+        }
+
+    def _unpack_missing_predictions(
+        self,
+        missing_predictions: NDArray[np.int32],
+        label_key_idx: int,
+        number_of_labels: int,
+        number_of_examples: int,
+    ) -> dict[
+        str,
+        dict[
+            str,
+            int | list[dict[str, str | tuple[float, float, float, float]]],
+        ],
+    ]:
+        """
+        Unpacks a numpy array of missing prediction counts and examples.
+        """
+
+        datum_idx = (
+            lambda gt_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+                missing_predictions[
+                    gt_label_idx,
+                    example_idx * 2 + 1,
+                ]
+            )
+        )
+
+        groundtruth_idx = (
+            lambda gt_label_idx, example_idx: int(  # noqa: E731 - lambda fn
+                missing_predictions[
+                    gt_label_idx,
+                    example_idx * 2 + 2,
+                ]
+            )
+        )
+
+        return {
+            self.index_to_label[gt_label_idx][1]: {
+                "count": max(
+                    int(missing_predictions[gt_label_idx, 0]),
+                    0,
+                ),
+                "examples": [
+                    {
+                        "datum": self.index_to_uid[
+                            datum_idx(gt_label_idx, example_idx)
+                        ],
+                        "groundtruth": tuple(
+                            self.groundtruth_examples[
+                                datum_idx(gt_label_idx, example_idx)
+                            ][
+                                groundtruth_idx(gt_label_idx, example_idx)
+                            ].tolist()
+                        ),
+                    }
+                    for example_idx in range(number_of_examples)
+                    if datum_idx(gt_label_idx, example_idx) >= 0
+                ],
+            }
+            for gt_label_idx in range(number_of_labels)
+            if (
+                self.label_index_to_label_key_index[gt_label_idx]
+                == label_key_idx
+            )
+        }
+
+    def _compute_confusion_matrix(
         self,
+        data: NDArray[np.floating],
+        label_metadata: NDArray[np.int32],
         iou_thresholds: list[float] = [0.5],
         score_thresholds: list[float] = [
             score / 10.0 for score in range(1, 11)
         ],
-        n_samples: int = 0,
-    ) -> list[DetailedCounts]:
+        number_of_examples: int = 0,
+    ) -> list[ConfusionMatrix]:
         """
         Computes detailed counting metrics.
 
@@ -462,132 +845,63 @@ class Evaluator:
             List of IoU thresholds to compute metrics for.
         score_thresholds : list[float], default=[0.1,0.2,...,1.0]
             List of confidence thresholds to compute metrics for.
-        n_samples : int, default=0
-            Number of datum samples to return per metric.
+        number_of_examples : int, default=0
+            Maximum number of annotation examples to return per metric.
 
         Returns
         -------
-        list[list[DetailedCounts]]
+        list[list[ConfusionMatrix]]
             Outer list is indexed by label, inner list is by IoU.
         """
 
-        if self._detailed_pairs.size == 0:
+        if data.size == 0:
             return list()
 
-        metrics = compute_detailed_counts(
-            self._detailed_pairs,
-            label_metadata=self._label_metadata,
+        (
+            confusion_matrix,
+            hallucinations,
+            missing_predictions,
+        ) = compute_confusion_matrix(
+            data=data,
+            label_metadata=label_metadata,
             iou_thresholds=np.array(iou_thresholds),
             score_thresholds=np.array(score_thresholds),
-            n_samples=n_samples,
+            n_examples=number_of_examples,
         )
 
-        tp_idx = 0
-        fp_misclf_idx = 2 * n_samples + 1
-        fp_halluc_idx = 4 * n_samples + 2
-        fn_misclf_idx = 6 * n_samples + 3
-        fn_misprd_idx = 8 * n_samples + 4
-
-        def _unpack_examples(
-            iou_idx: int,
-            label_idx: int,
-            type_idx: int,
-            example_source: dict[int, NDArray[np.float16]],
-        ) -> list[list[tuple[str, tuple[float, float, float, float]]]]:
-            """
-            Unpacks metric examples from computation.
-            """
-            type_idx += 1
-
-            results = list()
-            for score_idx in range(n_scores):
-                examples = list()
-                for example_idx in range(n_samples):
-                    datum_idx = metrics[
-                        iou_idx,
-                        score_idx,
-                        label_idx,
-                        type_idx + example_idx * 2,
-                    ]
-                    annotation_idx = metrics[
-                        iou_idx,
-                        score_idx,
-                        label_idx,
-                        type_idx + example_idx * 2 + 1,
-                    ]
-                    if datum_idx >= 0:
-                        examples.append(
-                            (
-                                self.index_to_uid[datum_idx],
-                                tuple(
-                                    example_source[datum_idx][
-                                        annotation_idx
-                                    ].tolist()
-                                ),
-                            )
-                        )
-                results.append(examples)
-
-            return results
-
-        n_ious, n_scores, n_labels, _ = metrics.shape
+        n_ious, n_scores, n_labels, _, _ = confusion_matrix.shape
         return [
-            DetailedCounts(
+            ConfusionMatrix(
                 iou_threshold=iou_thresholds[iou_idx],
-                label=self.index_to_label[label_idx],
-                score_thresholds=score_thresholds,
-                tp=metrics[iou_idx, :, label_idx, tp_idx].astype(int).tolist(),
-                fp_misclassification=metrics[
-                    iou_idx, :, label_idx, fp_misclf_idx
-                ]
-                .astype(int)
-                .tolist(),
-                fp_hallucination=metrics[iou_idx, :, label_idx, fp_halluc_idx]
-                .astype(int)
-                .tolist(),
-                fn_misclassification=metrics[
-                    iou_idx, :, label_idx, fn_misclf_idx
-                ]
-                .astype(int)
-                .tolist(),
-                fn_missing_prediction=metrics[
-                    iou_idx, :, label_idx, fn_misprd_idx
-                ]
-                .astype(int)
-                .tolist(),
-                tp_examples=_unpack_examples(
-                    iou_idx=iou_idx,
-                    label_idx=label_idx,
-                    type_idx=tp_idx,
-                    example_source=self.prediction_examples,
-                ),
-                fp_misclassification_examples=_unpack_examples(
-                    iou_idx=iou_idx,
-                    label_idx=label_idx,
-                    type_idx=fp_misclf_idx,
-                    example_source=self.prediction_examples,
-                ),
-                fp_hallucination_examples=_unpack_examples(
-                    iou_idx=iou_idx,
-                    label_idx=label_idx,
-                    type_idx=fp_halluc_idx,
-                    example_source=self.prediction_examples,
+                score_threshold=score_thresholds[score_idx],
+                label_key=label_key,
+                number_of_examples=number_of_examples,
+                confusion_matrix=self._unpack_confusion_matrix(
+                    confusion_matrix=confusion_matrix[
+                        iou_idx, score_idx, :, :, :
+                    ],
+                    label_key_idx=label_key_idx,
+                    number_of_labels=n_labels,
+                    number_of_examples=number_of_examples,
                 ),
-                fn_misclassification_examples=_unpack_examples(
-                    iou_idx=iou_idx,
-                    label_idx=label_idx,
-                    type_idx=fn_misclf_idx,
-                    example_source=self.groundtruth_examples,
+                hallucinations=self._unpack_hallucinations(
+                    hallucinations=hallucinations[iou_idx, score_idx, :, :],
+                    label_key_idx=label_key_idx,
+                    number_of_labels=n_labels,
+                    number_of_examples=number_of_examples,
                 ),
-                fn_missing_prediction_examples=_unpack_examples(
-                    iou_idx=iou_idx,
-                    label_idx=label_idx,
-                    type_idx=fn_misprd_idx,
-                    example_source=self.groundtruth_examples,
+                missing_predictions=self._unpack_missing_predictions(
+                    missing_predictions=missing_predictions[
+                        iou_idx, score_idx, :, :
+                    ],
+                    label_key_idx=label_key_idx,
+                    number_of_labels=n_labels,
+                    number_of_examples=number_of_examples,
                 ),
             )
-            for label_idx in range(n_labels)
+            for label_key_idx, label_key in self.index_to_label_key.items()
             for iou_idx in range(n_ious)
+            for score_idx in range(n_scores)
         ]
 
 
@@ -661,12 +975,27 @@ class DataLoader:
             self._evaluator.label_key_to_index[label[0]],
         )
 
-    def _add_data(
+    def _compute_ious_and_cache_pairs(
         self,
         uid_index: int,
         keyed_groundtruths: dict,
         keyed_predictions: dict,
-    ):
+        annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask],
+    ) -> None:
+        """
+        Compute IOUs between groundtruths and preditions before storing as pairs.
+
+        Parameters
+        ----------
+        uid_index: int
+            The index of the detection.
+        keyed_groundtruths: dict
+            A dictionary of groundtruths.
+        keyed_predictions: dict
+            A dictionary of predictions.
+        annotation_type: type[BoundingBox] | type[Polygon] | type[Bitmask]
+            The type of annotation to compute IOUs for.
+        """
         gt_keys = set(keyed_groundtruths.keys())
         pd_keys = set(keyed_predictions.keys())
         joint_keys = gt_keys.intersection(pd_keys)
@@ -677,14 +1006,13 @@ class DataLoader:
         for key in joint_keys:
             n_predictions = len(keyed_predictions[key])
             n_groundtruths = len(keyed_groundtruths[key])
-            boxes = np.array(
-                [
-                    np.array([*gextrema, *pextrema])
-                    for _, _, _, pextrema in keyed_predictions[key]
-                    for _, _, gextrema in keyed_groundtruths[key]
-                ]
+            data = _get_annotation_data(
+                keyed_groundtruths=keyed_groundtruths,
+                keyed_predictions=keyed_predictions,
+                key=key,
+                annotation_type=annotation_type,
             )
-            ious = compute_iou(boxes)
+            ious = compute_iou(data=data, annotation_type=annotation_type)
             mask_nonzero_iou = (ious > 1e-9).reshape(
                 (n_predictions, n_groundtruths)
             )
@@ -780,9 +1108,10 @@ class DataLoader:
 
         self.pairs.append(np.array(pairs))
 
-    def add_data(
+    def _add_data(
         self,
         detections: list[Detection],
+        annotation_type: type[Bitmask] | type[BoundingBox] | type[Polygon],
         show_progress: bool = False,
     ):
         """
@@ -792,6 +1121,8 @@ class DataLoader:
         ----------
         detections : list[Detection]
             A list of Detection objects.
+        annotation_type : type[Bitmask] | type[BoundingBox] | type[Polygon]
+            The annotation type to process.
         show_progress : bool, default=False
             Toggle for tqdm progress bar.
         """
@@ -817,45 +1148,146 @@ class DataLoader:
             # cache labels and annotations
             keyed_groundtruths = defaultdict(list)
             keyed_predictions = defaultdict(list)
+
+            representation_property = _get_annotation_representation(
+                annotation_type=annotation_type
+            )
+
             for gidx, gann in enumerate(detection.groundtruths):
-                self._evaluator.groundtruth_examples[uid_index][
-                    gidx
-                ] = np.array(gann.extrema)
+                if not isinstance(gann, annotation_type):
+                    raise ValueError(
+                        f"Expected {annotation_type}, but annotation is of type {type(gann)}."
+                    )
+
+                if isinstance(gann, BoundingBox):
+                    self._evaluator.groundtruth_examples[uid_index][
+                        gidx
+                    ] = getattr(gann, representation_property)
+                else:
+                    converted_box = gann.to_box()
+                    self._evaluator.groundtruth_examples[uid_index][gidx] = (
+                        getattr(converted_box, "extrema")
+                        if converted_box is not None
+                        else None
+                    )
                 for glabel in gann.labels:
                     label_idx, label_key_idx = self._add_label(glabel)
                     self.groundtruth_count[label_idx][uid_index] += 1
+                    representation = getattr(gann, representation_property)
                     keyed_groundtruths[label_key_idx].append(
                         (
                             gidx,
                             label_idx,
-                            gann.extrema,
+                            representation,
                         )
                     )
+
             for pidx, pann in enumerate(detection.predictions):
-                self._evaluator.prediction_examples[uid_index][
-                    pidx
-                ] = np.array(pann.extrema)
+                if not isinstance(pann, annotation_type):
+                    raise ValueError(
+                        f"Expected {annotation_type}, but annotation is of type {type(pann)}."
+                    )
+
+                if isinstance(pann, BoundingBox):
+                    self._evaluator.prediction_examples[uid_index][
+                        pidx
+                    ] = getattr(pann, representation_property)
+                else:
+                    converted_box = pann.to_box()
+                    self._evaluator.prediction_examples[uid_index][pidx] = (
+                        getattr(converted_box, "extrema")
+                        if converted_box is not None
+                        else None
+                    )
                 for plabel, pscore in zip(pann.labels, pann.scores):
                     label_idx, label_key_idx = self._add_label(plabel)
                     self.prediction_count[label_idx][uid_index] += 1
+                    representation = representation = getattr(
+                        pann, representation_property
+                    )
                     keyed_predictions[label_key_idx].append(
                         (
                             pidx,
                             label_idx,
                             pscore,
-                            pann.extrema,
+                            representation,
                         )
                     )
 
-            self._add_data(
+            self._compute_ious_and_cache_pairs(
                 uid_index=uid_index,
                 keyed_groundtruths=keyed_groundtruths,
                 keyed_predictions=keyed_predictions,
+                annotation_type=annotation_type,
             )
 
-    def add_data_from_valor_dict(
+    def add_bounding_boxes(
+        self,
+        detections: list[Detection],
+        show_progress: bool = False,
+    ):
+        """
+        Adds bounding box detections to the cache.
+
+        Parameters
+        ----------
+        detections : list[Detection]
+            A list of Detection objects.
+        show_progress : bool, default=False
+            Toggle for tqdm progress bar.
+        """
+        return self._add_data(
+            detections=detections,
+            show_progress=show_progress,
+            annotation_type=BoundingBox,
+        )
+
+    def add_polygons(
+        self,
+        detections: list[Detection],
+        show_progress: bool = False,
+    ):
+        """
+        Adds polygon detections to the cache.
+
+        Parameters
+        ----------
+        detections : list[Detection]
+            A list of Detection objects.
+        show_progress : bool, default=False
+            Toggle for tqdm progress bar.
+        """
+        return self._add_data(
+            detections=detections,
+            show_progress=show_progress,
+            annotation_type=Polygon,
+        )
+
+    def add_bitmasks(
+        self,
+        detections: list[Detection],
+        show_progress: bool = False,
+    ):
+        """
+        Adds bitmask detections to the cache.
+
+        Parameters
+        ----------
+        detections : list[Detection]
+            A list of Detection objects.
+        show_progress : bool, default=False
+            Toggle for tqdm progress bar.
+        """
+        return self._add_data(
+            detections=detections,
+            show_progress=show_progress,
+            annotation_type=Bitmask,
+        )
+
+    def _add_data_from_valor_dict(
         self,
         detections: list[tuple[dict, dict]],
+        annotation_type: type[Bitmask] | type[BoundingBox] | type[Polygon],
         show_progress: bool = False,
     ):
         """
@@ -865,20 +1297,14 @@ class DataLoader:
         ----------
         detections : list[tuple[dict, dict]]
             A list of groundtruth, prediction pairs in Valor-format dictionaries.
+        annotation_type : type[Bitmask] | type[BoundingBox] | type[Polygon]
+            The annotation type to process.
         show_progress : bool, default=False
             Toggle for tqdm progress bar.
         """
 
-        def _get_bbox_extrema(
-            data: list[list[list[float]]],
-        ) -> tuple[float, float, float, float]:
-            x = [point[0] for shape in data for point in shape]
-            y = [point[1] for shape in data for point in shape]
-            return (min(x), max(x), min(y), max(y))
-
         disable_tqdm = not show_progress
         for groundtruth, prediction in tqdm(detections, disable=disable_tqdm):
-
             # update metadata
             self._evaluator.n_datums += 1
             self._evaluator.n_groundtruths += len(groundtruth["annotations"])
@@ -898,10 +1324,34 @@ class DataLoader:
             # cache labels and annotations
             keyed_groundtruths = defaultdict(list)
             keyed_predictions = defaultdict(list)
+
+            annotation_key = _get_valor_dict_annotation_key(
+                annotation_type=annotation_type
+            )
+            invalid_keys = list(
+                filter(
+                    lambda x: x != annotation_key,
+                    ["bounding_box", "raster", "polygon"],
+                )
+            )
+
             for gidx, gann in enumerate(groundtruth["annotations"]):
-                self._evaluator.groundtruth_examples[uid_index][
-                    gidx
-                ] = np.array(_get_bbox_extrema(gann["bounding_box"]))
+                if (gann[annotation_key] is None) or any(
+                    [gann[k] is not None for k in invalid_keys]
+                ):
+                    raise ValueError(
+                        f"Input JSON doesn't contain {annotation_type} data, or contains data for multiple annotation types."
+                    )
+                if annotation_type == BoundingBox:
+                    self._evaluator.groundtruth_examples[uid_index][
+                        gidx
+                    ] = np.array(
+                        _get_annotation_representation_from_valor_dict(
+                            gann[annotation_key],
+                            annotation_type=annotation_type,
+                        ),
+                    )
+
                 for valor_label in gann["labels"]:
                     glabel = (valor_label["key"], valor_label["value"])
                     label_idx, label_key_idx = self._add_label(glabel)
@@ -910,13 +1360,29 @@ class DataLoader:
                         (
                             gidx,
                             label_idx,
-                            _get_bbox_extrema(gann["bounding_box"]),
+                            _get_annotation_representation_from_valor_dict(
+                                gann[annotation_key],
+                                annotation_type=annotation_type,
+                            ),
                         )
                     )
             for pidx, pann in enumerate(prediction["annotations"]):
-                self._evaluator.prediction_examples[uid_index][
-                    pidx
-                ] = np.array(_get_bbox_extrema(pann["bounding_box"]))
+                if (pann[annotation_key] is None) or any(
+                    [pann[k] is not None for k in invalid_keys]
+                ):
+                    raise ValueError(
+                        f"Input JSON doesn't contain {annotation_type} data, or contains data for multiple annotation types."
+                    )
+
+                if annotation_type == BoundingBox:
+                    self._evaluator.prediction_examples[uid_index][
+                        pidx
+                    ] = np.array(
+                        _get_annotation_representation_from_valor_dict(
+                            pann[annotation_key],
+                            annotation_type=annotation_type,
+                        )
+                    )
                 for valor_label in pann["labels"]:
                     plabel = (valor_label["key"], valor_label["value"])
                     pscore = valor_label["score"]
@@ -927,16 +1393,41 @@ class DataLoader:
                             pidx,
                             label_idx,
                             pscore,
-                            _get_bbox_extrema(pann["bounding_box"]),
+                            _get_annotation_representation_from_valor_dict(
+                                pann[annotation_key],
+                                annotation_type=annotation_type,
+                            ),
                         )
                     )
 
-            self._add_data(
+            self._compute_ious_and_cache_pairs(
                 uid_index=uid_index,
                 keyed_groundtruths=keyed_groundtruths,
                 keyed_predictions=keyed_predictions,
+                annotation_type=annotation_type,
             )
 
+    def add_bounding_boxes_from_valor_dict(
+        self,
+        detections: list[tuple[dict, dict]],
+        show_progress: bool = False,
+    ):
+        """
+        Adds Valor-format bounding box detections to the cache.
+
+        Parameters
+        ----------
+        detections : list[tuple[dict, dict]]
+            A list of groundtruth, prediction pairs in Valor-format dictionaries.
+        show_progress : bool, default=False
+            Toggle for tqdm progress bar.
+        """
+        return self._add_data_from_valor_dict(
+            detections=detections,
+            show_progress=show_progress,
+            annotation_type=BoundingBox,
+        )
+
     def finalize(self) -> Evaluator:
         """
         Performs data finalization and some preprocessing steps.