**7.2 Neighborhood suppression results**

The suppression of neighboring positions was tested on the standard frontal face MIT+CMU dataset. Three WaldBoost classifiers with target false positive rates of 0.01, 0.05 and 0.2 were trained for four types of image features: LRD, LRP, LBP and Haar. For each classifier, three neighborhood suppression strategies were trained with target false positive rates of 0.01, 0.05 and 0.2. Comparing results of the combinations allows us to evaluate if it is more effective to use neighborhood suppression than just by using a WaldBoost classifier with a higher false positive rate. The results of this experiment in Fig. 8 clearly show that neighborhood suppression is indeed effective and on average it evaluates less weak hypotheses per image position for the same accuracy.

### **7.3 EnMS results**

EnMS was evaluated on a face localization task. The dataset was downloaded from Flicker groups *portraits* (training) and *just\_faces* (testing). The dataset contains 84, 251 training and 6, 704 near-frontal faces. The images were rescaled to a 100 × 100 pixel resolution with the face approximately 50 × 50 pixels large and positioned in the middle. Both WaldBoost and

EnMS were evaluated on this data for several target false negative rates. The localization accuracy was measured as the number of images where the detector returned a position with the highest response of a classifier which always evaluated all weak hypotheses. In order to allow for some tolerance, errors were also counted as failure to detect position with the reference classifier response lower by 2 and 6 than the best response and failure to detect position with the reference response higher than 2 which is an operating point that still gives reasonably low false alarms in the detection task. The results in Fig. 9 show that EnMS provides approximately two times better speed for the same error rates than WaldBoost.

Real-Time Algorithms of Object Detection Using Classifiers 245

This chapter focused on methods of real-time object detection with classifiers. It has been demonstrated that the object detection methods working in real-time are feasible and can be implemented on a variety of platforms, such as personal computer processors, GP-GPU

In order to achieve real-time performance, an efficient implementation platform and efficient implementation itself is necessary, but further enhancement through algorithmic acceleration is needed as well. Two examples of such acceleration are presented in the chapter: exploitation of information about neighborhoods of the already classified positions in the image and early suppression of non-maxima of the classifier responses. The approach of exploitation of the neighborhoods in the image is based on the idea that classification of the overlapping sub-images in the image - the neighborhoods - may share some properties and information. One of the possible ways to share such information is through re-using the weak classifiers used during classification of one location through WaldBoost for predicting results in the other neighboring locations. This prediction is done through a machine learning process similar to WaldBoost where the difference to WaldBoost is that the training process actually reuses the already selected weak classifiers that were used at the original location. While this process

Pre-processing that rules out some parts of the image from the detection process can significantly speed up the detection process. Important future research certainly includes machine-learning based pre-processing methods and research of under-sampling in scanning methods that can also improve detection performance possibly without any adverse effects on precision. Future research also includes algorithmic improvements of acceleration methods, such as improvement in the processor assignment in GP-GPU, improved scanning trajectories

This work has been supported by the FIT VUT Brno project "Advanced recognition and presentation of multimedia data", FIT VUT, FIT-S-11-2, Centre of excellence in computer science "The IT4Innovations Centre of Excellence", EU, CZ 1.05/1.1.00/02.0070, "Reduced Certification Costs Using Trusted Multi-core Platforms", Artemis JU, RECOMP #100202, "Smart Multicore Embedded Systems", Artemis JU, SMECY #100230, and the Czech Ministry of Education, Youth and Sports, "Security-Oriented Research in Information Technology",

CEZ MŠMT, MSM0021630528 and "Centre of Computer Graphics", MŠMT, LC06008.

works well only in close neighborhoods, it brings a significant speed-up.

in neighborhood exploitation, or further improvements in feature extraction.

**8. Conclusions**

**9. Acknowledgement**

platforms, or even in programmable hardware.

Fig. 8. The graphs show AUC on y-axis (Area Under ROC) versus the average number of weak classifiers evaluated per image position as measured on the MIT+CMU frontal face dataset. The individual lines are for original WaldBoost detectors without neighborhood suppression (full line) and the other lines are with added neighborhood suppression with different target false negative rates. Good results should be in the left (fast) bottom (accurate) corner.

Fig. 9. The graphs show frontal face localization error (y-axis) for different speed-ups achieved by WaldBoost and EnMS. The speed-up is measured as reduction of the number of weak hypotheses evaluated on average per image position relative to the full length of the classifier (length is the same for WaldBoost and EnMS). The lines represent differently computed errors (see text) of WaldBoost and EnMS.

EnMS were evaluated on this data for several target false negative rates. The localization accuracy was measured as the number of images where the detector returned a position with the highest response of a classifier which always evaluated all weak hypotheses. In order to allow for some tolerance, errors were also counted as failure to detect position with the reference classifier response lower by 2 and 6 than the best response and failure to detect position with the reference response higher than 2 which is an operating point that still gives reasonably low false alarms in the detection task. The results in Fig. 9 show that EnMS provides approximately two times better speed for the same error rates than WaldBoost.
