File size: 7,379 Bytes
97b6013 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 |
# Supported object detection evaluation protocols
The Tensorflow Object Detection API currently supports three evaluation protocols,
that can be configured in `EvalConfig` by setting `metrics_set` to the
corresponding value.
## PASCAL VOC 2010 detection metric
`EvalConfig.metrics_set='pascal_voc_detection_metrics'`
The commonly used mAP metric for evaluating the quality of object detectors,
computed according to the protocol of the PASCAL VOC Challenge 2010-2012. The
protocol is available
[here](http://host.robots.ox.ac.uk/pascal/VOC/voc2010/devkit_doc_08-May-2010.pdf).
## Weighted PASCAL VOC detection metric
`EvalConfig.metrics_set='weighted_pascal_voc_detection_metrics'`
The weighted PASCAL metric computes the mean average precision as the average
precision when treating all classes as a single class. In comparison,
PASCAL metrics computes the mean average precision as the mean of the
per-class average precisions.
For example, the test set consists of two classes, "cat" and "dog", and there
are ten times more boxes of "cat" than those of "dog". According to PASCAL VOC
2010 metric, performance on each of the two classes would contribute equally
towards the final mAP value, while for the Weighted PASCAL VOC metric the final
mAP value will be influenced by frequency of each class.
## PASCAL VOC 2010 instance segmentation metric
`EvalConfig.metrics_set='pascal_voc_instance_segmentation_metrics'`
Similar to Pascal VOC 2010 detection metric, but computes the intersection over
union based on the object masks instead of object boxes.
## Weighted PASCAL VOC instance segmentation metric
`EvalConfig.metrics_set='weighted_pascal_voc_instance_segmentation_metrics'`
Similar to the weighted pascal voc 2010 detection metric, but computes the
intersection over union based on the object masks instead of object boxes.
## COCO detection metrics
`EvalConfig.metrics_set='coco_detection_metrics'`
The COCO metrics are the official detection metrics used to score the
[COCO competition](http://cocodataset.org/) and are similar to Pascal VOC
metrics but have a slightly different implementation and report additional
statistics such as mAP at IOU thresholds of .5:.95, and precision/recall
statistics for small, medium, and large objects.
See the
[pycocotools](https://github.com/cocodataset/cocoapi/tree/master/PythonAPI)
repository for more details.
## COCO mask metrics
`EvalConfig.metrics_set='coco_mask_metrics'`
Similar to the COCO detection metrics, but computes the
intersection over union based on the object masks instead of object boxes.
## Open Images V2 detection metric
`EvalConfig.metrics_set='oid_V2_detection_metrics'`
This metric is defined originally for evaluating detector performance on [Open
Images V2 dataset](https://github.com/openimages/dataset) and is fairly similar
to the PASCAL VOC 2010 metric mentioned above. It computes interpolated average
precision (AP) for each class and averages it among all classes (mAP).
The difference to the PASCAL VOC 2010 metric is the following: Open Images
annotations contain `group-of` ground-truth boxes (see [Open Images data
description](https://github.com/openimages/dataset#annotations-human-bboxcsv)),
that are treated differently for the purpose of deciding whether detections are
"true positives", "ignored", "false positives". Here we define these three
cases:
A detection is a "true positive" if there is a non-group-of ground-truth box,
such that:
* The detection box and the ground-truth box are of the same class, and
intersection-over-union (IoU) between the detection box and the ground-truth
box is greater than the IoU threshold (default value 0.5). \
Illustration of handling non-group-of boxes: \
* yellow box - ground-truth box;
* green box - true positive;
* red boxes - false positives.
* This is the highest scoring detection for this ground truth box that
satisfies the criteria above.
A detection is "ignored" if it is not a true positive, and there is a `group-of`
ground-truth box such that:
* The detection box and the ground-truth box are of the same class, and the
area of intersection between the detection box and the ground-truth box
divided by the area of the detection is greater than 0.5. This is intended
to measure whether the detection box is approximately inside the group-of
ground-truth box. \
Illustration of handling `group-of` boxes: \
* yellow box - ground-truth box;
* grey boxes - two detections on cars, that are ignored;
* red box - false positive.
A detection is a "false positive" if it is neither a "true positive" nor
"ignored".
Precision and recall are defined as:
* Precision = number-of-true-positives/(number-of-true-positives + number-of-false-positives)
* Recall = number-of-true-positives/number-of-non-group-of-boxes
Note that detections ignored as firing on a `group-of` ground-truth box do not
contribute to the number of true positives.
The labels in Open Images are organized in a
[hierarchy](https://storage.googleapis.com/openimages/2017_07/bbox_labels_vis/bbox_labels_vis.html).
Ground-truth bounding-boxes are annotated with the most specific class available
in the hierarchy. For example, "car" has two children "limousine" and "van". Any
other kind of car is annotated as "car" (for example, a sedan). Given this
convention, the evaluation software treats all classes independently, ignoring
the hierarchy. To achieve high performance values, object detectors should
output bounding-boxes labelled in the same manner.
The old metric name is DEPRECATED.
`EvalConfig.metrics_set='open_images_V2_detection_metrics'`
## OID Challenge Object Detection Metric
`EvalConfig.metrics_set='oid_challenge_detection_metrics'`
The metric for the OID Challenge Object Detection Metric 2018/2019 Object
Detection track. The description is provided on the
[Open Images Challenge website](https://storage.googleapis.com/openimages/web/evaluation.html#object_detection_eval).
The old metric name is DEPRECATED.
`EvalConfig.metrics_set='oid_challenge_object_detection_metrics'`
## OID Challenge Visual Relationship Detection Metric
The metric for the OID Challenge Visual Relationship Detection Metric 2018,2019
Visual Relationship Detection track. The description is provided on the
[Open Images Challenge website](https://storage.googleapis.com/openimages/web/evaluation.html#visual_relationships_eval).
Note: this is currently a stand-alone metric, that can be used only through the
`metrics/oid_vrd_challenge_evaluation.py` util.
## OID Challenge Instance Segmentation Metric
`EvalConfig.metrics_set='oid_challenge_segmentation_metrics'`
The metric for the OID Challenge Instance Segmentation Metric 2019, Instance
Segmentation track. The description is provided on the
[Open Images Challenge website](https://storage.googleapis.com/openimages/web/evaluation.html#instance_segmentation_eval).
|