rpx_benchmark.metrics.detection
Object detection + open-vocab detection metric calculators.
1"""Object detection + open-vocab detection metric calculators.""" 2 3from __future__ import annotations 4 5from typing import Dict 6 7from ..api import DetectionGroundTruth, DetectionPrediction, TaskType 8from ..exceptions import MetricError 9from .registry import MetricCalculator, register_metric 10 11 12@register_metric(TaskType.OBJECT_DETECTION) 13@register_metric(TaskType.OPEN_VOCAB_DETECTION) 14class DetectionMetrics(MetricCalculator): 15 """Precision / recall / F1 at a single IoU threshold. 16 17 Parameters 18 ---------- 19 iou_threshold : float 20 Minimum IoU (0.0–1.0) for a prediction to count as a true 21 positive. Defaults to 0.5, the COCO baseline. 22 23 Notes 24 ----- 25 Greedy matching by descending prediction score. A GT box is 26 consumed after the first predicted box matches it, so subsequent 27 predictions for the same object register as false positives 28 (standard VOC/COCO evaluation rule). 29 """ 30 31 name = "detection_metrics" 32 33 def __init__(self, iou_threshold: float = 0.5) -> None: 34 self.iou_threshold = iou_threshold 35 36 def compute( 37 self, 38 prediction: DetectionPrediction, 39 ground_truth: DetectionGroundTruth, 40 ) -> Dict[str, float]: 41 if not isinstance(prediction, DetectionPrediction): 42 raise MetricError( 43 f"DetectionMetrics expected DetectionPrediction, got {type(prediction).__name__}", 44 ) 45 if not isinstance(ground_truth, DetectionGroundTruth): 46 raise MetricError( 47 f"DetectionMetrics expected DetectionGroundTruth, got " 48 f"{type(ground_truth).__name__}", 49 ) 50 51 from ..evaluators import detection_metrics 52 53 return detection_metrics( 54 pred_boxes=prediction.boxes, 55 pred_scores=prediction.scores, 56 pred_labels=prediction.labels, 57 gt_boxes=ground_truth.boxes, 58 gt_labels=ground_truth.labels, 59 iou_threshold=self.iou_threshold, 60 )
@register_metric(TaskType.OBJECT_DETECTION)
@register_metric(TaskType.OPEN_VOCAB_DETECTION)
class
DetectionMetrics13@register_metric(TaskType.OBJECT_DETECTION) 14@register_metric(TaskType.OPEN_VOCAB_DETECTION) 15class DetectionMetrics(MetricCalculator): 16 """Precision / recall / F1 at a single IoU threshold. 17 18 Parameters 19 ---------- 20 iou_threshold : float 21 Minimum IoU (0.0–1.0) for a prediction to count as a true 22 positive. Defaults to 0.5, the COCO baseline. 23 24 Notes 25 ----- 26 Greedy matching by descending prediction score. A GT box is 27 consumed after the first predicted box matches it, so subsequent 28 predictions for the same object register as false positives 29 (standard VOC/COCO evaluation rule). 30 """ 31 32 name = "detection_metrics" 33 34 def __init__(self, iou_threshold: float = 0.5) -> None: 35 self.iou_threshold = iou_threshold 36 37 def compute( 38 self, 39 prediction: DetectionPrediction, 40 ground_truth: DetectionGroundTruth, 41 ) -> Dict[str, float]: 42 if not isinstance(prediction, DetectionPrediction): 43 raise MetricError( 44 f"DetectionMetrics expected DetectionPrediction, got {type(prediction).__name__}", 45 ) 46 if not isinstance(ground_truth, DetectionGroundTruth): 47 raise MetricError( 48 f"DetectionMetrics expected DetectionGroundTruth, got " 49 f"{type(ground_truth).__name__}", 50 ) 51 52 from ..evaluators import detection_metrics 53 54 return detection_metrics( 55 pred_boxes=prediction.boxes, 56 pred_scores=prediction.scores, 57 pred_labels=prediction.labels, 58 gt_boxes=ground_truth.boxes, 59 gt_labels=ground_truth.labels, 60 iou_threshold=self.iou_threshold, 61 )
Precision / recall / F1 at a single IoU threshold.
Parameters
iou_threshold : float Minimum IoU (0.0–1.0) for a prediction to count as a true positive. Defaults to 0.5, the COCO baseline.
Notes
Greedy matching by descending prediction score. A GT box is consumed after the first predicted box matches it, so subsequent predictions for the same object register as false positives (standard VOC/COCO evaluation rule).
def
compute( self, prediction: rpx_benchmark.api.DetectionPrediction, ground_truth: rpx_benchmark.api.DetectionGroundTruth) -> Dict[str, float]:
37 def compute( 38 self, 39 prediction: DetectionPrediction, 40 ground_truth: DetectionGroundTruth, 41 ) -> Dict[str, float]: 42 if not isinstance(prediction, DetectionPrediction): 43 raise MetricError( 44 f"DetectionMetrics expected DetectionPrediction, got {type(prediction).__name__}", 45 ) 46 if not isinstance(ground_truth, DetectionGroundTruth): 47 raise MetricError( 48 f"DetectionMetrics expected DetectionGroundTruth, got " 49 f"{type(ground_truth).__name__}", 50 ) 51 52 from ..evaluators import detection_metrics 53 54 return detection_metrics( 55 pred_boxes=prediction.boxes, 56 pred_scores=prediction.scores, 57 pred_labels=prediction.labels, 58 gt_boxes=ground_truth.boxes, 59 gt_labels=ground_truth.labels, 60 iou_threshold=self.iou_threshold, 61 )
Return a dict of metric name → scalar for one sample.
Parameters
prediction : Any
Task-specific Prediction dataclass
(e.g. DepthPrediction). The concrete type is
whatever the calculator is designed for.
ground_truth : Any
Task-specific GroundTruth dataclass
(e.g. DepthGroundTruth).
Returns
dict[str, float] Metric values. Keys should be short snake_case names; values must be numeric so the runner can average them.
Raises
MetricError When inputs have the wrong shape/type/content.