Home

Awesome

mAP (mean Average Precision)

GitHub stars

This code will evaluate the performance of your neural net for object recognition.

<p align="center"> <img src="https://user-images.githubusercontent.com/15831541/37559643-6738bcc8-2a21-11e8-8a07-ed836f19c5d9.gif" width="450" height="300" /> </p>

In practice, a higher mAP value indicates a better performance of your neural net, given your ground-truth and set of classes.

Citation

This project was developed for the following paper, please consider citing it:

@INPROCEEDINGS{8594067,
  author={J. {Cartucho} and R. {Ventura} and M. {Veloso}},
  booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={Robust Object Recognition Through Symbiotic Deep Learning In Mobile Robots}, 
  year={2018},
  pages={2336-2341},
}

Table of contents

Explanation

The performance of your neural net will be judged using the mAP criterium defined in the PASCAL VOC 2012 competition. We simply adapted the official Matlab code into Python (in our tests they both give the same results).

First (1.), we calculate the Average Precision (AP), for each of the classes present in the ground-truth. Finally (2.), we calculate the mAP (mean Average Precision) value.

1. Calculate AP

For each class:

First, your neural net detection-results are sorted by decreasing confidence and are assigned to ground-truth objects. We have "a match" when they share the same label and an IoU >= 0.5 (Intersection over Union greater than 50%). This "match" is considered a true positive if that ground-truth object has not been already used (to avoid multiple detections of the same object).

<img src="https://user-images.githubusercontent.com/15831541/37725175-45b9e1a6-2d2a-11e8-8c15-2fb4d716ca9a.png" width="35%" height="35%" />

Using this criterium, we calculate the precision/recall curve. E.g:

<img src="https://user-images.githubusercontent.com/15831541/43008995-64dd53ce-8c34-11e8-8a2c-4567b1311910.png" width="45%" height="45%" />

Then we compute a version of the measured precision/recall curve with precision monotonically decreasing (shown in light red), by setting the precision for recall r to the maximum precision obtained for any recall r' > r.

Finally, we compute the AP as the area under this curve (shown in light blue) by numerical integration. No approximation is involved since the curve is piecewise constant.

2. Calculate mAP

We calculate the mean of all the AP's, resulting in an mAP value from 0 to 100%. E.g:

<img src="https://user-images.githubusercontent.com/15831541/38933241-5f9556ae-4310-11e8-9d47-cb205f9b103b.png"/> <img src="https://user-images.githubusercontent.com/15831541/38933180-366b6fca-4310-11e8-99b9-17ad4b159b86.png" />

Prerequisites

You need to install:

Optional:

Quick-start

To start using the mAP you need to clone the repo:

git clone https://github.com/Cartucho/mAP

Running the code

Step by step:

  1. Create the ground-truth files
  2. Copy the ground-truth files into the folder input/ground-truth/
  3. Create the detection-results files
  4. Copy the detection-results files into the folder input/detection-results/
  5. Run the code: python main.py

Optional (if you want to see the animation):

  1. Insert the images into the folder input/images-optional/

PASCAL VOC, Darkflow and YOLO users

In the scripts/extra folder you can find additional scripts to convert PASCAL VOC, darkflow and YOLO files into the required format.

Create the ground-truth files

Create the detection-results files

Authors: