[email protected] Explained: What mAP50 Measures and How to Read It in Practice

mAP50 is the class mean of average precision at IoU 0.5. How it is calculated, how it differs from mAP50-95, and why a high score can mislead.

mAP@0.5 Explained: What mAP50 Measures and How to Read It in Practice
Written by TechnoLynx Published on 11 Jul 2026

mAP50 (also written [email protected]) is the mean, across object classes, of average precision: the area under each class’s precision-recall curve, where a predicted box counts as correct only if it overlaps a same-class ground-truth box with an intersection over union (IoU) of 0.5 or more. A mAP50 of 0.71 therefore means the detector ranks its correct detections well when “correct” is defined loosely, and does not reflect how tightly those boxes localise the objects. That gap between “found the object” and “boxed it accurately” is the main thing to keep in mind when you read the number.

How mAP50 is calculated from precision and recall at IoU 0.5

Everything starts with IoU, the area of overlap between a predicted box and a ground-truth box divided by the area of their union, which runs from 0 (no overlap) to 1 (perfect match). At the 0.5 threshold, a prediction becomes a true positive when its IoU with a not-yet-matched label of the same class reaches 0.5; anything else is a false positive. A ground-truth object that no prediction matches is a false negative.

From those counts come the two familiar ratios. Precision is true positives divided by all predicted boxes, and recall is true positives divided by all ground-truth boxes.

Average precision is computed from a precision-recall curve rather than a single pair of numbers. Evaluation pools every detection of one class across the whole dataset, sorts them by confidence, and walks down the list: each detection either claims the best-overlapping unmatched label and becomes a true positive, or becomes a false positive. A label can be claimed only once, so a second box on an object that has already been found counts against precision, which is the PASCAL VOC rule for duplicates. Precision and recall are recomputed after every step, tracing the curve from the most confident detection to the least, and AP is the area under that curve. mAP50 is the plain average of those per-class AP values.

If you want the wider vocabulary, our guide to object detection metrics for inspection covers it.

Worked example: one class, five detections, two thresholds

The numbers below are a constructed toy case to show the mechanism, not a measurement. Assume one class, four ground-truth objects, and five detections already sorted by confidence.

Rank Best IoU with an unmatched label At IoU 0.5 Precision Recall At IoU 0.75
1 0.82 TP 1/1 = 1.00 1/4 = 0.25 TP
2 0.71 TP 2/2 = 1.00 2/4 = 0.50 FP
3 0.55 FP 2/3 = 0.67 0.50 FP
4 0.35 TP 3/4 = 0.75 3/4 = 0.75 FP
5 lowest (duplicate of object 1) FP 3/5 = 0.60 0.75 FP

The fourth object is never detected, so recall tops out at 0.75. Using all-point interpolation, where precision at each recall level is replaced by the highest precision achieved at that recall or above, the precision envelope is 1.00 up to recall 0.50 and 0.75 from 0.50 to 0.75. AP at IoU 0.5 is then 0.50 × 1.00 + 0.25 × 0.75 = 0.6875.

Raise the threshold to 0.75 and only the first detection survives as a true positive, so AP falls to 0.25 × 1.00 = 0.25. The detector did not change; only the definition of “correct” did.

Why two tools give different mAP50 for the same predictions

AP is computed from the curve at sampled recall points, and the sampling differs between benchmarks. The PASCAL VOC 2007 challenge integrated over 11 equally spaced recall points, while MS COCO uses 101, with precision interpolated to the maximum obtained at higher recall. Later PASCAL VOC implementations moved to interpolating all data points rather than 11. A survey of detection literature describes the 11-step version as taking the maximum precision at recall values from 0 to 1 in steps of 0.1 and averaging them.

mAP50 vs mAP50-95: which number to report

mAP50-95 averages mAP over ten IoU thresholds, 0.50 to 0.95 in steps of 0.05, and COCO treats that average as its primary challenge metric, calling it simply “AP”. COCO’s own description is explicit about the motive: averaging over IoUs rewards detectors with better localisation, and computing AP at a single IoU of 0.5 is described as the traditional approach it breaks from. COCO labels AP at IoU 0.50 “the PASCAL VOC metric” and AP at IoU 0.75 “the strict metric”.

Because a looser threshold counts more detections as correct, mAP50 always comes out highest, mAP50-95 below it. Published pairs show how wide that gap runs in practice:

Detector and dataset (source) mAP50 mAP50-95 mAP50-95 ÷ mAP50
YOLOv2, DarkNet-19, COCO (detection survey, 2018) 44.0 21.6 0.49
Faster R-CNN with FPN, ResNet-101, COCO (same survey) 59.1 36.2 0.61
RetinaNet, ResNeXt-101-FPN, COCO (same survey) 61.1 40.8 0.67
Top-scoring entry (Thousandwords), ODOR olfactory object challenge 18.93% 11.49% 0.61
YOLO26 nano, federated training, KIIT-MiTA drone data (2026 study) 0.7137 0.4695 0.66

The ratio column is simple division of the published figures. The YOLOv2 row is the instructive one: it keeps less than half of its AP50 once tighter boxes are demanded, which matches the survey’s remark that YOLO reports better accuracy on AP at 0.5 than on the full mAP metric.

Which to report depends on what the boxes are for:

Your situation Lead with Why
Counting, presence/absence, triggering a downstream step mAP50 Box tightness beyond rough overlap does not change the outcome
Measuring, grasping, cropping for a second model, precision-critical robotics or medical work mAP50-95, with mAP75 alongside Practitioners commonly optimise for mAP50:95 in precision-critical applications
Comparing against published COCO results mAP50-95 It is the number COCO winners are ranked on
Labels known to be loose or inconsistent mAP50, flagged as such PASCAL VOC set 0.5 deliberately low to allow for inaccurate ground-truth boxes

Our view is that reporting both, plus mAP75, removes most of the ambiguity, and Ultralytics validation already emits all three. For a deeper treatment of reading the gap as a localisation signal, see our comparison of the two metrics for localisation quality.

Why a model can show a high mAP50 but a low mAP50-95

A large gap is a localisation signal, and YOLOv3 is one documented case. Its authors reported 57.9 AP50 in 51 ms on a Titan X against 57.5 AP50 in 198 ms for RetinaNet, yet wrote that “performance drops significantly as the IOU threshold increases indicating YOLOv3 struggles to get the boxes perfectly aligned with the object”.

COCO’s error analysis of a ResNet detector shows how much localisation can account for. At IoU 0.75 its AP was 0.399; perfect localisation would have lifted it to 0.682, while removing all class confusions would only have raised it to 0.713, and the analysis concluded its errors were dominated by imperfect localisation and background confusion. The TIDE error-analysis toolbox (Bolya et al., ECCV 2020) found the same pattern from the other side: localisation error is the error type that responds most strongly as the IoU threshold rises, so the 0.5:0.95 average is heavily biased toward localisation errors. TIDE also notes that Mask Scoring R-CNN made no significant difference at AP50 but gained substantially at higher thresholds, a change you could not see by watching AP50 alone.

The looser threshold has defenders, and their argument is fair. Redmon and Farhadi cite Russakovsky et al. reporting that humans find it surprisingly hard to tell an IoU of 0.3 from 0.5, and argue that COCO’s emphasis on box quality must de-emphasise something else, in their reading classification accuracy. Our reading is that both sides are right about different applications: if nobody downstream uses the exact box edges, a mAP50-95 penalty for slightly loose boxes is measuring something you do not need.

mAP50 is high but precision and recall are low: what does that mean?

mAP50 summarises the whole confidence-ranked precision-recall curve, whereas the precision and recall a tool prints are values at one operating point, so the two can tell different stories without either being wrong. The evaluation views detection as ranking detected instances per category. A detector that puts its correct detections at the top of the list earns a high AP even if, further down, it emits a long tail of wrong boxes: the Optimal Correction Cost paper (2022) shows a detector producing many incorrect detections that mAP does not penalise because they rank below the correct one.

No source we used measures this mismatch directly, so what follows is reasoning from the definition. Three causes are worth separating:

  • The operating threshold is in the wrong place. If the reported precision and recall come from a confidence cut-off far from the knee of the curve, they understate what the ranking supports; picking a better threshold fixes it without retraining.
  • The low-confidence tail is noisy. mAP tends to encourage detectors to output many low-confidence detections, because its priority is raising the chance of true positives rather than penalising false ones. Precision measured with that tail included will look poor.
  • Per-class imbalance. mAP is an unweighted mean of class APs, so a few easy, high-AP classes can lift the headline while precision and recall pooled over many instances of a hard class stay low.

Choosing that operating point is a separate decision from the metric, and our article on confidence scores in computer vision covers how to set one.

Is a mAP50 of 0.5 good?

There is no universal threshold: a mAP50 of 0.5 can be excellent on one dataset and weak on another. The published figures above make the point. The winning entry in the ODOR olfactory object challenge scored 18.93% [email protected] on its test set, while the YOLO26 nano model in the 2026 drone study reached 0.7137 (71.37%), and these numbers come from different datasets with different task difficulty rather than a single ladder. Object size alone moves the result a long way: in the ODOR challenge, every submission struggled with small boxes, and the winner’s mAP fell from 11.71% on medium boxes to 4.19% on small ones.

Annotation quality caps the ceiling too. TIDE found that 30 of the 100 most confident localisation errors in Mask R-CNN detections on COCO came from bad annotations, rising to 50 of 100 for background errors, and that misannotated ground truth considerably lowers the maximum mAP a reasonable model can achieve.

The useful question is therefore relative. We would judge a mAP50 against a baseline trained and evaluated on the same data with the same evaluator, against the per-class APs rather than only the mean, and against what the application needs at its chosen operating point.

Reading mAP50 in Ultralytics YOLO validation output

Ultralytics validation reports mAP50, mAP75 and mAP50-95 for detection models. In the Python API, metrics.box.map50 is mAP50, metrics.box.map75 is mAP75, metrics.box.map is mAP50-95, and metrics.box.maps is a list of mAP50-95 values per category. Note the naming trap: the attribute called plain map is the strict COCO-style average, not mAP50.

Ultralytics documents that the minimum confidence threshold for validation is used by the precision-recall curves, mAP and the confusion matrices, and that lower values increase recall but may add false positives. Because that threshold feeds the precision-recall curves and mAP, compare runs only at the same conf setting. Per-image precision, recall, F1, TP, FP and FN are stored at IoU 0.5, and with save_json=True detection validation also reports small-, medium- and large-object mAP. Our explainer on how mAP scores YOLO detectors goes further into the YOLO-specific output.

Does a high validation mAP50 predict production performance?

Not on its own: detectors are trained and evaluated on data from the same source distribution, while deployment data often differ, leading to substantial performance degradation. Al-Emadi et al. (CVPR 2025) note that detectors are typically trained and evaluated under the i.i.d. assumption, and that in real deployment target distributions often differ. Image corruption alone degrades detectors: Michaelis et al. (2019) showed a range of standard detectors falling to 30 to 60% of their original performance on corrupted images, while stylising the training images substantially improved robustness.

The metric also has blind spots that matter more in deployment than on a leaderboard. mAP does not treat images equally, and a detector is not penalised for producing no detections at all in some images. Whether a given false positive affects mAP depends on detections in other images. And the Optimal Correction Cost authors observed that mAP distributions overlap across detectors, so the performance ranking of detectors can be flipped. A line-side camera that produces no detections in some images is not penalised for those images by mAP.

For that reason we treat validation mAP50 as an entry ticket, not a release criterion. The checks aimed at these blind spots are an evaluation set built from deployment conditions (lighting, camera, weather, rare classes), per-image error checks alongside the dataset-level mean, and monitoring once the model is live; these are the checks we would build into computer vision R&D work and the drift side of MLOps engineering. The artefacts that keep a deployed detector honest are covered in industrial CV inspection production reliability, and for safety-reviewed systems in the automotive perception validation package. Edge cases that aggregate scores smooth over have their own treatment in why benchmark accuracy hides perception edge cases, and where detections feed a tracker, multiple object tracking in production covers where that chain breaks.

What mAP50 is good for, stated narrowly

mAP50 answers one question well: across the dataset, does the detector rank roughly-right boxes of each class above wrong ones? It does not answer how tight the boxes are, what happens at your operating threshold, whether every image is covered, or how the model behaves once the camera, the weather or the product mix changes. Read it next to mAP75 and mAP50-95, the per-class list and the size breakdown, on an evaluation set that looks like deployment. The question still open for any given project is which of those failure modes your application can tolerate, and that is decided by the downstream use of the boxes, not by the metric.

Back See Blogs
arrow icon