Real-time Distracted Driver Posture Classification
`Yehya Abouelnaga
`Department of Informatics
`Technical University of Munich
`yehya.abouelnaga@tum.de
`Hesham M. Eraqi
`Department of Computer Science and Engineering
`The American University in Cairo
`heraqi@aucegypt.edu
`Mohamed N. Moustafa
`Department of Computer Science and Engineering
`The American University in Cairo
`m.moustafa@aucegypt.edu
`Abstract
`In this paper, we present a new dataset for “distracted driver” posture estimation.
`In addition, we propose a novel system that achieves 95.98% driving posture
`estimation classification accuracy. The system consists of a genetically-weighted
`ensemble of Convolutional Neural Networks (CNNs). We show that a weighted
`ensemble of classifiers using a genetic algorithm yields in better classification
`confidence. We also study the effect of different visual elements (i.e. hands and face)
`in distraction detection and classification by means of face and hand localizations.
`Finally, we present a thinned version of our ensemble that could achieve a 94.29%
`classification accuracy and operate in a realtime environment.
`1 Introduction
`The number of road accidents due to distracted driving is steadily increasing. According to the
`National Highway Traffic Safety Administration (NHTSA), in 2015, 3,477 people were killed, and
`391,000 were injured in motor vehicle crashes involving distracted drivers Pickrell et al. [2016].
`The major cause of these accidents was the use of mobile phones. The NHTSA defines distracted
`driving as “any activity that diverts attention from driving”, including: a) Talking or Texting on one’s
`phone, b) eating and drinking, c) talking to passengers, d) fiddling with the stereo, entertainment,
`or navigation system Pickrell et al. [2016]. The Center for Disease Control and Prevention (CDC)
`provides a broader definition of distracted driving by taking into account visual (i.e. taking one’s eyes
`off the road), manual (i.e. taking one’s hands off the driving wheel) and cognitive (i.e. taking one’s
`mind off driving) causes Services [2016]. We believe that the detection of distracted driver’s postures
`is key to further preventive measures. Distracted driver detection is also important for autonomous
`vehicles; Latest commercial self-driving cars still require drivers to pay attention and be ready to take
`back control of the wheel Eriksson and Stanton [2017].
`We present a realtime distracted driver pose estimation system using a weighted ensemble of con-
`volutional neural networks and a challenging distracted driver’s dataset on which we evaluate our
`proposed solution.
`2 Literature Review
`The work in the distracted driver detection field over the past seven years could be clustered into four
`groups: multiple independent cell phone detection publications, Laboratory of Intelligent and Safe
`Automobiles in University of California San Diego (UCSD) datasets and publications, Southeast
`32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
`arXiv:1706.09498v3 [cs.CV] 29 Nov 2018
`Samsara EX1032
`Samsara v. Motive
`IPR2026-00108
`
`
`
`
`
`
`
`University Distracted Driver dataset and affiliated publications, and recently, StateFarm’s Distracted
`Driver Kaggle competition.
`2.1 Cell Phone Usage Detection
`Berri and Silva [2014] presents an SVM-based model that detects the use of mobile phone while
`driving (i.e. distracted driving). Their dataset consists of frontal image view of a driver’s face. They
`also make pre-made assumptions about hand and face locations in the picture. Craye and Karray
`[2015] uses AdaBoost classifier and Hidden Markov Models to classify a Kinect’s RGB-D data.
`Their solution depends on data produced by indoor data. They sit on a chair and a mimmic a certain
`distraction (i.e. talking on the phone). This setup misses two essential points: the lighting conditions
`and the distance between a Kinect and the driver. In real-life applications, a driver is exposed to
`a variety of lighting conditions (i.e. sunlight and shadow). Hoang Ngan Le et al. [2016] devised
`a Faster-RCNN model to detect driver’s cell-phone usage and “hands on the wheel”. Their model
`is mainly geared towards face/hand segmentation. They train their Faster-RCNN on the dataset
`proposed in Das et al. [2015] (that we also use in this paper). Their proposed solution runs at a 0.06,
`and 0.09 frames per second for cell-phone usage, and “hands on the wheel” detection.
`2.2 UCSD’s Laboratory of Intelligent and Safe Automobiles Work
`In Ohn-bar and Martin [2013], the authors present a fusion of classifiers where they segment the
`image to three regions: wheel, gear, and instrument panel (i.e. radio). They develop a classifier for
`each segment in which they detect existence of hands in those areas. The information from these
`scenes are passed to an “activity classifier” that detects the actual activity (i.e. adjusting the radio,
`operating the gear). Ohn-bar and Trivedi [2013a] presents a region-based classification approach. It
`detects hands presence in certain pre-defined regions in an image. A model is learned for each region
`separately. All regions are later joined using a second-stage classifier. Ohn-bar and Trivedi [2013b]
`collects a new RGBD dataset in which they observe the driving wheel and a driver’s hand activity.
`The frames are divided into 5 labelled regions with classes: One hand, no hands, two hands, two
`hands + cell, two hands + map, and two hands + bottle.
`2.3 Southeast University Distracted Driver Dataset
`Zhao et al. [2011a] designed a more inclusive distracted driving dataset with a side view of the driver
`and more activities: grasping the steering wheel, operating the shift lever, eating a cake and talking
`on a cellular phone. In their paper, they introduced a contourlet transform for feature extraction, and
`then, evaluated the performance of different classifiers: Random Forests (RF), k-nearest neighbors
`classifier (KNN), and Multi-Layer Perceptron (MLP) classifier. The random forests achieved the
`highest classification accuracy of 90.5%. Zhao et al. [2012] showed that using a multiwavelet
`transform improves the accuracy of multilayer perceptron classifier to 90.61% (previously 37.06%).
`Zhao et al. [2013] improves the Multilayer Perceptron (MLP) classifier using combined features of
`Pyramid Histogram of Oriented Gradients (PHOG) and spatial scale feature extractors. Their MLP
`achieves a 94.75% classification accuracy. Yan et al. [2016a] introduces a R*CNN that trains on
`manually labelled pre-defined regions (i.e. driver, shift lever). Their convolutional nerual net achieves
`a 97.76%. It is worth noting that all previous publications tested their accuracies against four classes.
`This publication tested against six classes. Yan et al. [2016b] presents a convolutional neural network
`solution that achieves a 99.78% classification accuracy. They train their network in a 2-step process.
`First, they use pre-trained sparse filters as the parameters of the first convolutional layer. Second, they
`fine-tune the network on the actuall dataset. Their accuracy is measured against the 4-classes of the
`Southeast dataset: wheel (safe driving), eating/smoking, operating the shift lever, and talking on the
`phone.
`2.4 StateFarm’s Dataset
`StateFarm’s Distracted Driver Detection competition on Kaggle was the first publicly available dataset
`for posture classification. In the competition, StateFarm defined ten postures to be detected: safe
`driving, texting using right hand, talking on the phone using right hand, texting using left hand, talking
`on the phone using left hand, operating the radio, drinking, reaching behind, hair and makeup, and
`talking to passenger. Our work, in this paper, is mainly inspired by StateFarm’s Distracted Driver’s
`2
`
`
`
`
`
`
`
`competition. While the usage of StateFarm’s dataset is limited to the purposes of the competition
`Sultan [2016], we designed a similar dataset that follows the same postures.
`3 Dataset Design
`Figure 1: Examples of the American University in Cairo (AUC) Distracted Driver’s Dataset. In a
`column-level order, postures are: drinking, adjusting the radio, driving in a safe posture, fiddling with
`hair or makeup, reaching behind, talking to passengers, talk on cell phone using left hand, talk on cell
`phone using right hand, texting using left hand, and texting using right hand.
`Creating a new dataset (“AUC Distracted Driver” dataset) was essential to the completion of this
`work. The available alternatives to our dataset are: StateFarm and Southeast University (SEU)
`datasets. StateFarm’s dataset is to be used for their Kaggle past competition purpose only (as
`per their regulations) Sultan [2016]. As per our multiple attempts to obtain it, we knew that the
`authors of Southeast University (SEU) dataset do not make it publicly available. Also, their dataset
`consists of only four postures. All the papers (Yan et al. [2016a,b, 2014], Zhao et al. [2013, 2012,
`2011b,a]) that benchmarked against the dataset are affiliated with the either Southeast University,
`Xi’an Jiaotong-Liverpool University, or Liverpool University, and they have at least one shared author.
`The dataset was collected using an ASUS ZenPhone (Model Z00UD) rear camera. The input was
`collected in a video format, and then, cut into individual images, 1080 × 1920 each. The phone was
`fixed using an arm strap to the car roof handle on top of the passenger’s seat. In our use case, this
`setup proved to be very flexible as we needed to collect data in different vehicles. In order to label
`the collected videos, we designed a simple multi-platform action annotation tool. The annotation tool
`is open-source and publicly available at Abouelnaga [2017].
`We had 31 participants from 7 different countries: Egypt (24), Germany (2), USA (1), Canada (1),
`Uganda (1), Palestine (1), and Morocco (1). Out of all participants, 22 were males and 9 were females.
`Videos were shot in 4 different cars: Proton Gen2 (26), Mitsubishi Lancer (2), Nissan Sunny (2), and
`KIA Carens (1).
`4 Proposed Method
`Our proposed solution consists of a genetically-weighted ensemble of convolutional neural networks.
`The convolutional neural networks train on raw images, face images, hands images, and “face+hands”
`images. We train an AlexNet Krizhevsky et al. [2012] and an InceptionV3 Szegedy et al. [2016] on
`those four images sources. In the InceptionV3 network, we fine-tune a pre-trained ImageNet model
`(i.e. transfer learning). Then, we evaluate a weighted sum of all networks’ outputs yielding the final
`class distribution. The weights are evaluated using a genetic algorithm.
`4.1 Face & Hands Detection
`We trained the model presented in Li, Haoxiang and Lin, Zhe and Shen, Xiaohui and Brandt, Jonathan
`and Hua [2015] on the Annotated Facial Landmarks in the Wild (AFLW) face dataset Koestinger
`et al. [2011]. The trained model achieved decent results. However, it was sensitive to distance from
`the camera (i.e. faces that were close to the camera were not easily detected). We found that the
`pre-trained model (presented in Farfade et al. [2015]) produced better results on our dataset. Given
`that we did not have any hand labelled face bounding boxes, we couldn’t formally compare the two
`models. However, when randomly selecting images from different classes, we found that Farfade
`et al. [2015] was closer to what we expected.
`3
`
`
`
`
`
`
`
`Figure 2: An overview of our proposed solution. A face detector and a hand detector are run against
`each frame. For each output image (i.e. Face and Hands), an AlexNet and an InceptionV3 networks
`are trained (i.e. resulting in 8 neural networks: 4 AlexNet and 4 InceptionV3). The overall class
`distribution is determined by the weighted sum of all softmax layers. The weights are learned using a
`genetic algorithm.
`As for hands detection, we used the pre-trained model presented in Bambach et al. [2015] with slight
`modifications. Their trained model was a binary class AlexNet that classifies hands/non-hands for
`different proposal windows. We transferred the weights of the fully connected layers (i.e. fc6, fc7 and
`fc8) into convolutional layers such that each neuron in the fully connected layer was transferred into
`a depth layer with a 1-pixel kernel size. Except the first fully connected layer. Also, this architecture
`accepts variant size inputs and produces variant-size outputs. The last convolutional layer has a
`depth of 2 (i.e. the binary classes) where Conv8x,y,0 + Conv8x,y,1 = 1 for all x and y; such that
`0 ≤ x < W, 0 ≤ y < Hand W and H are the output’s width and height, respectively.
`4.2 Convolutional Neural Network
`For distracted driver posture classification, we trained two classes of neural networks: AlexNet
`and InceptionV3. Each network is trained on 4 different image sources (i.e. raw, face, hands and
`face+hands images) yielding in 4 models per net and a total of 8 models.
`We trained our AlexNet models from scratch. We didn’t use a pre-trained model. For InceptionV3,
`we performed a transfer learning. We fine-tuned a pre-trained model on the distraction postures. We
`removed the “logits” layer, and replaced it with a 10-neuron fully connected layer (i.e. corresponding
`to 10 driving postures).
`We used a gradient descent optimizer with an initial learning rate of 10−2. The learning rate decays
`linearly in each epoch with a step of (10−2 −10−4)/Epochs. We trained the networks for 30 epochs.
`In each, we divide the training dataset into mini-batches of 50 images each.
`4.3 Weighted Ensemble of Classifiers using Genetic Algorithm
`Each classifier produces a class probability vector (i.e. output of “softmax” layer), C1 . . . CN , such
`that Ci has 10 probabilities (i.e. 10 classes) andN is the number of classifiers (N = 8in our situation).
`In a majority voting system, we assume that all experts (i.e. classifiers) can equally contribute to a
`better decision by taking the unweighted sum of all classifier outputs.
`CMajority = 1
`N
`N∑
`i
`Ci, C Weighted = 1∑N
`i wi
`N∑
`i
`wi · Ci
`However, that is not usually a valid assumption. In a weighted voting system, we assume that
`classifiers do not contribute equally to the ensemble and that some classifiers might yield higher
`accuracy than others. Therefore, we need to estimate the weights of each classifier’s contribution to
`the ensemble. Rokach [2010] presents a variety of methods to estimate the weights. We opted to use
`a genetic algorithm (i.e. a search-based method).
`4
`
`
`
`
`
`
`
`Table 1: Distracted Driver Posture Classification Results
`Model Source Loss (NLL) Accuracy (%)
`AlexNet
`Original 0.3909 93.65
`Face 1.0516 84.28
`Hands 0.6186 89.52
`Face + Hands 0.8298 86.68
`InceptionV3
`Original 0.2654 95.17
`Face 0.6096 88.82
`Hands 0.4546 91.62
`Face + Hands 0.4495 90.88
`Realtime System 0.2727 94.29
`Majority Voting Ensemble 0.1661 95.77
`GA-Weighted Ensemble 0.1575 95.98
`Our chromosome consists ofN genes that correspond to the weightsw1 . . . wN . Our fitness function
`evaluates the Negative Log Likelihood (NLL) loss over a 50% random sample of the population. This
`helps prevent overfitting. Our population consists of 50 individual. In each iteration, we retain the
`top 20% of the population and use them as parents. Then, we randomly select 10% of the remaining
`80% of the population as parents. In other words, we have 30% of the population as parents. Now,
`we randomly mutate 5% of the selected parents. Finally, we cross-over random pairs of the parents
`to produce children until we have a full population (i.e. with 50 individuals). We ran the above
`procedure for only 5 iterations in order to avoid over-fitting. We selected the chromosome with the
`highest fitness score (test against all data points– not 50%).
`5 Experiments
`We divided our dataset into 75% training and 25% held out test data. Then, we ran the face and hand
`detectors on the entire dataset. We tested all of the networks against our test dataset and obtained
`the results in Table 1. We notice that both AlexNet and InceptionV3 achieve best accuracies when
`trained on the original images. Hands seem to have more weight in posture recognition than the
`face. “Face + Hands” images produce slightly lower accuracy than the hands images, yet, still higher
`than the face images. That happens due to face/hand detector failures. For example, if a hand is not
`found, we pass a face image to a “face + hands” classifier. This doesn’t happen in individual cases of
`hand-only or face-only classifier because if the hand/face detection fails, we pass the original image
`to the hand/face classifier as a fallback mechanism. With better hand/face detectors, the “face+hands”
`networks are expected to produce higher accuracies than the “hands” networks. An ensemble of
`two AlexNet models produce a satisfactory classification accuracy (i.e. 94.29%). Meanwhile, it still
`maintains a realtime performance on a CPU-based system.
`We trained and tested our models using an EVGA GeForce GTX TITAN X 12GB GPU, Intel(R)
`Core(TM) i7-5960X CPU @ 3.00GHz, and a 48 GM RAM. On average, AlexNet processed 182
`frames per second using a GPU and 52 frames per second using a CPU. InceptionV3 processes 72
`frames per second using a GPU and 5.5 frames per second using a CPU.
`5.1 Analysis
`In Table 2, we notice that the most confusing posture is the “safe driving”. This is due to the lack
`of temporal context in static images. In a static image, a driver would appear in a “safe driving”
`posture. However, contextually, he/she was distracted by doing some other activity. “Text Left” is
`mostly confused for “Talk Left” and vice versa. Same applies to “Text Right” and “Talk Right”.
`“Adjust Radio” is mainly confused for a “safe driving” posture. That is due to lack of the previously
`5
`
`
`
`
`
`
`
`Table 2: Confusion Matrix of Genetically Weighted Ensemble of Classifiers
`Predicted
`C0 C1 C2 C3 C4 C5 C6 C7 C8 C9
`Actual
`C0 95.34 0 0.33 0.65 0.11 0.43 0.43 0.87 0.11 1.74
`C1 0.31 96.63 1.23 0.31 0.92 0 0.31 0 0.31 0
`C2 0.29 3.23 96.48 0 0 0 0 0 0 0
`C3 2.02 0.61 0 96.15 0.81 0 0.20 0 0 0.20
`C4 0 0.33 0 4.90 94.77 0 0 0 0 0
`C5 4.26 0 0 0.33 0 95.08 0 0 0 0.33
`C6 0.74 0 0 0.25 0 0.74 98.01 0.25 0 0
`C7 3.65 0 0 0 0 0 0 95.35 0 1.00
`C8 3.79 0 0 0 0 0 1.38 0.34 92.76 1.72
`C9 1.40 0 0 0 0 0 0.47 0.31 0.16 97.67
`mentioned temporal context. Apart from safe driving, “Hair & Makeup” is confused for talking to
`passenger. That is because, in most cases, when drivers did their hair/makeup on the left side of
`their face, they needed to tilt their face slightly right (while looking at the frontal mirror). Thus,
`the network thought the person was talking to passenger. “Reach Behind” was confused for both
`talking to passenger and drinking. That makes sense as people tend to naturally look towards the
`camera while reaching behind. As for the drinking confusion, it is due to right-arm movement from
`the steering wheel to the back seat. A still image in the middle of that move could be easily mistaken
`for a drinking posture. “Drink” and “Talk to Passenger” postures were not easily confused with other
`postures as 98% and 97.67% of their images were correctly classified.
`6 Conclusion
`Distracted driving is a major problem leading to a striking number of accidents worldwide. In
`addition to regulatory measures to tackle such problems, we believe that smart vehicles would indeed
`contribute to a safer driving experience. In this paper, we presented a robust vision-based system
`that recognizes distracted driving postures. We collected a challenging distracted driver dataset that
`we used to develop and test our system. Our best model utilizes a genetically weighted ensemble of
`convolutional neural networks to achieve a 95.98% accuracy. We also showed that a simpler model
`(only using AlexNet) could operate in realtime and still maintain a satisfactory classification accuracy.
`Face and hands detection is proved to improve classification accuracy in our ensemble. However, in a
`realtime setting, their performance overhead is much higher than their contribution.
`In a future work, we need to devise a better face and hands detector. We would need to manually
`label hand and face proposals and use them to train an object detector (i.e. SSD) to improve faces and
`hands localization. In order to overcome the “safe driving” posture confusion with other classes, we
`would need to incorporate temporality in our decision. We shall test the performance of a Recurrent
`Neural Network (RNN) against sequential stream of frames. We envision a performance improvement
`due to temporal features.
`References
`Timothy M. Pickrell, Hongying (Ruby) Li, and Shova KC. TRAFFIC SAFETY FACTS, 2016. URL
`https://www.nhtsa.gov/risky-driving/distracted-driving.
`US Department of Health & Human Services. Distracted Driving, 2016. URL https://www.cdc.
`gov/motorvehiclesafety/distracted{_}driving/.
`6
`
`
`
`
`
`
`
`Alexander Eriksson and Neville A Stanton. Takeover time in highly automated vehicles: Noncritical
`transitions to and from manual control. Human factors, 59(4):689–705, 2017.
`Rafael Berri and Alexandre Gonçalves Silva. A Pattern Recognition System for Detecting Use of
`Mobile Phones While Driving. (August), 2014. doi: 10.5220/0004684504110418.
`Céline Craye and Fakhri Karray. Driver distraction detection and recognition using RGB-D sensor.
`arXiv preprint arXiv:1502.00250, 2015.
`T Hoang Ngan Le, Yutong Zheng, Chenchen Zhu, Khoa Luu, and Marios Savvides. Multiple Scale
`Faster-RCNN Approach to Driver’s Cell-Phone Usage and Hands on Steering Wheel Detection.
`Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops,
`pages 46–53, 2016. ISSN 21607516. doi: 10.1109/CVPRW.2016.13.
`Nikhil Das, Eshed Ohn-bar, and Mohan M Trivedi. On Performance Evaluation of Driver Hand
`Detection Algorithms : Challenges , Dataset , and Metrics. 2015.
`Eshed Ohn-bar and Sujitha Martin. Driver hand activity analysis in naturalistic driving studies :
`challenges , algorithms , and experimental studies challenges , algorithms , and experimental
`studies. Journal of Electronic Imaging, 22(4), 2013. doi: 10.1117/1.JEI.22.4.041119.
`Eshed Ohn-bar and Mohan Trivedi. In-Vehicle Hand Activity Recognition Using Integration of
`Regions. Intelligent Vehicles Symposium (IV), 2013 IEEE, pages 1034—-1039, 2013a.
`Eshed Ohn-bar and Mohan M Trivedi. The Power is in Your Hands : 3D Analysis of Hand Gestures in
`Naturalistic Video. 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops,
`pages 912—-917, 2013b. doi: 10.1109/CVPRW.2013.134.
`C H Zhao, B L Zhang, J He, and J Lian. Recognition of driving postures by contourlet transform and
`random forests. IET Intelligent Transport Systems, 6(2):161–168, 2011a.
`Chihang Zhao, Yongsheng Gao, Jie He, and Jie Lian. Recognition of driving postures by multiwavelet
`transform and multilayer perceptron classifier. Engineering Applications of Artificial Intelligence,
`25(8):1677–1686, 2012.
`Chihang H Zhao, Bailing L Zhang, Xiaozheng Z Zhang, Sanqiang Q Zhao, and Hanxi X Li. Recog-
`nition of driving postures by combined features and random subspace ensemble of multilayer
`perceptron classifiers. Neural Computing and Applications, 22(1):175–184, 2013.
`Shiyang Yan, Yuxuan Teng, Jeremy S. Smith, and Bailing Zhang. Driver behavior recognition based
`on deep convolutional neural networks. 12th International Conference on Natural Computation,
`Fuzzy Systems and Knowledge Discovery (ICNC-FSKD), pages 636—-641, 2016a. doi: 10.1109/
`FSKD.2016.7603248. URL http://ieeexplore.ieee.org/document/7603248/.
`Chao Yan, Frans Coenen, and Bailing Zhang. Driving posture recognition by convolutional neural
`networks. IET Computer Vision, 10(2):103–114, 2016b. ISSN 1751-9632. doi: 10.1049/iet-cvi.
`2015.0175.
`Ihab Sultan. Academic purposes?, 2016. URL https://www.kaggle.com/c/
`state-farm-distracted-driver-detection/discussion/20043{#}114916 .
`Chao Yan, Frans Coenen, and Bailing Zhang. Driving Posture Recognition by Joint Application of
`Motion History Image and Pyramid Histogram of Oriented Gradients. International Journal of
`Vehicular Technology, 2014, 2014.
`Chihang Zhao, Bailing Zhang, Jie Lian, Jie He, Tao Lin, and Xiaoxiao Zhang. Classification of
`driving postures by support vector machines. Proceedings - 6th International Conference on Image
`and Graphics, ICIG 2011, (June 2014):926–930, 2011b. doi: 10.1109/ICIG.2011.184.
`Yehya Abouelnaga. Action Annotation Tool, 2017. URL https://github.com/devyhia/
`action-annotation.
`Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolu-
`tional neural networks. In Advances in neural information processing systems, pages 1097–1105,
`2012.
`7
`
`
`
`
`
`
`
`Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the
`inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer
`Vision and Pattern Recognition, pages 2818–2826, 2016.
`Gang Li, Haoxiang and Lin, Zhe and Shen, Xiaohui and Brandt, Jonathan and Hua. A Convolutional
`Neural Network Approach for Face Identification. Cvpr, pages 5325–5334, 2015. ISSN 1063-
`6919. doi: 10.1109/CVPR.2015.7299170. URL http://users.eecs.northwestern.edu/
`{~}xsh835/assets/cvpr2015{_}cascnn.pdf.
`Martin Koestinger, Paul Wohlhart, Peter M Roth, and Horst Bischof. Annotated Facial Landmarks in
`the Wild: A Large-scale, Real-world Database for Facial Landmark Localization. In First IEEE
`International Workshop on Benchmarking Facial Image Analysis Technologies, 2011.
`Sachin Sudhakar Farfade, Mohammad Saberian, and Li-jia Li. Multi-view Face Detection Using Deep
`Convolutional Neural Networks. Cornell University Library, page 37, 2015. ISSN 0920-5691.
`doi: 10.1007/s11263-015-0816-y. URL http://arxiv.org/abs/1409.0575{%}5Cnhttp:
`//arxiv.org/abs/1502.02766.
`Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending A Hand: Detecting Hands and
`Recognizing Activities in Complex Egocentric Interactions. In The IEEE International Conference
`on Computer Vision (ICCV), dec 2015.
`Lior Rokach. Ensemble-based classifiers. Artificial Intelligence Review, 33(1):1–39, 2010.
`8
`
`
`
`
`
`
`
`