Where is integration performed when computing mAP?

[LucFrachon]

Hi all!
I did a deep dive into the code for detection mAP and saw that the final average precision values are computed in MeanAveragePrecision._summarize() (full code below). As you can see, the code filters the precision values and simply computes the mean (second to last row).

But the AP is supposed to be the area under the PR curve, so each precision value should be weighted by the recall step size to approximate a function integration. I cannot see this done anywhere. What am I missing?

    def _summarize(
        self,
        results: Dict,
        avg_prec: bool = True,
        iou_threshold: Optional[float] = None,
        area_range: str = "all",
        max_dets: int = 100,
    ) -> Tensor:
        """Perform evaluation for single class and image.

        Args:
            results:
                Dictionary including precision, recall and scores for all combinations.
            avg_prec:
                Calculate average precision. Else calculate average recall.
            iou_threshold:
                IoU threshold. If set to ``None`` it all values are used. Else results are filtered.
            area_range:
                Bounding box area range key.
            max_dets:
                Maximum detections.
        """
        area_inds = [i for i, k in enumerate(self.bbox_area_ranges.keys()) if k == area_range]
        mdet_inds = [i for i, k in enumerate(self.max_detection_thresholds) if k == max_dets]
        if avg_prec:
            # dimension of precision: [TxRxKxAxM]
            prec = results["precision"]
            # IoU
            if iou_threshold is not None:
                thr = self.iou_thresholds.index(iou_threshold)
                prec = prec[thr, :, :, area_inds, mdet_inds]
            else:
                prec = prec[:, :, :, area_inds, mdet_inds]
        else:
            # dimension of recall: [TxKxAxM]
            prec = results["recall"]
            if iou_threshold is not None:
                thr = self.iou_thresholds.index(iou_threshold)
                prec = prec[thr, :, :, area_inds, mdet_inds]
            else:
                prec = prec[:, :, area_inds, mdet_inds]

        mean_prec = torch.tensor([-1.0]) if len(prec[prec > -1]) == 0 else torch.mean(prec[prec > -1])
        return mean_prec