Version: 2.0
OS: Linux (Ubuntu 14.04 is recommended)
- OpenCV2
- Python2
Follow the link to get multiple Haar classifier cascades.
- obj_detect.py: The function that you will use
- support/ The cpp code that supports the function of "obj_detect.py"
- cascade/ Trained classifiers. Include AT&T, Chase, Macy's, Mobil, Subway, Walmart, and banana (Taco Bell and Starbucks will come soon...).
- history/ Previous version of logo detection code
- images/ Images w/ or w/o logos
- tool/ Includes a script that can help you find the best MARGIN and THRESHOLD to have acceptable FPR and FNR. There is also an analysis result text file generated by me.
log/contains detection analysis log with the margin from 0 to 25plot_round_num/has the images that show the distribution of round number with different margins
Given image path and logo name, the "get_result" function in "obj_detect.py" will return the x-coordinate of the detected logo.
Put "obj_detect.py" in the same path with the "support" folder. Then import "obj_detect.py" into your code and call "get_result" function.
Sample: In python, use:
- $ res = get_result("walmart", "images/walmart/walmart.jpg", 1)
-
Input parameters:
- "walmart" is the name of logo that you want to detect
- "images/walmart/walmart.jpg" is the path of the image to process
- 1 means show the detection result; 0 means do not show the result (Displaying the result will block the get_result function and need PRESS_ANY_KEY to continue)
-
Returned string:
- If the program returns correctly, "res" will be in the format like "logo_left / logo_right / image_width / min_neighbor / conf_score".
- For "logo_left" and "logo_right", -1 means more than one object found; 0 means no object;
- If something goes wrong, "res" will be "ERROR"
In object detection, there is a score called "MIN_NEIGHBOR" which represents the similarity between the detected area and sample object. Moreover, I applied recursive image cropping to get the number of "cropping round", which can also be used to know if there is a real logo or not. In most cases, real logo will have higher round number while the false positive has less.
Threshold of conf_score To remove false positive cases, we need to ignore the results with confidence scores smaller than a certain threshold. The relation between the threshold, FPR, and FNR is partially listed below. You can select a threshold according to your application scenario.
- Threshold / FPR(%) / FNR(%)
- 10 / 15.9 / 13.0
- 11 / 15.9 / 14.4
- 12 / 15.9 / 14.4
- 13 / 14.4 / 14.4
- 14 / 13.0 / 14.4
- 15 / 11.5 / 15.9
- 16 / 11.5 / 15.9
- 17 / 10.1 / 18.8
- 18 / 8.6 / 18.8
- 19 / 8.6 / 20.2
- 20 / 5.7 / 21.7 (The data is tested among 64 images in 2015 May, new result will be updated shortly)
- Chase cascade updated (with the trainining set of streetview as background)
Please cite this paper in your publications if it helps your research
@inproceedings{Hu:2016:AAL:2971648.2971674,
author = {Hu, Yitao and Liu, Xiaochen and Nath, Suman and Govindan, Ramesh},
title = {ALPS: Accurate Landmark Positioning at City Scales},
booktitle = {Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing},
series = {UbiComp '16},
year = {2016},
isbn = {978-1-4503-4461-6},
location = {Heidelberg, Germany},
pages = {1147--1158},
numpages = {12},
url = {http://doi.acm.org/10.1145/2971648.2971674},
doi = {10.1145/2971648.2971674},
acmid = {2971674},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {context-aware computing, landmark localization system, machine/deep learning},
}