Keeping the safe distance from the leading vehicle is crucial for transportation companies with a fleet of old cars. While modern Advanced Driver Assistant Systems (ADAS) might be able to estimate the distance from the front-leading vehicle, traditional ADAS do not usually offer this feature. An alternative solution is to monitor the distance using smartphones that are attached to a place such as a sun visor. The basic idea behind this approach is to detect the front-leading vehicle using the smartphone camera and estimate its distance from the car. Although SSD can achieve real-time performance on powerful GPUs, it remains challenging to run this model in real-time on mobile devices. In this paper, we propose a monocular distance estimator for forward-leading vehicles using a smartphone which is faster and more accurate than the state-of-the-art SSD detector. Specifically, we propose a layer-wise method to generate more efficient default boxes for the SSD and develop a lightweight method for estimating the distance accurately. Our experiments show that the proposed method reduces the number of default boxes by an average of 38.4% while it improves the detection rate and the processing speed compared to the original SSD. Moreover, our monocular distance estimator provides a proper safety buffer zone when the distance is greater than 20 meters. A sample video is available at https://youtu.be/-ptvfabBZWA.

The cost of drawing object bounding boxes (i.e. labeling) for millions of images is prohibitively high. For instance, labeling pedestrians in a regular urban image could take 35 seconds on average. Active learning aims to reduce the cost of labeling by selecting only those images that are informative to improve the detection network accuracy. In this paper, we propose a method to perform active learning of object detectors based on convolutional neural networks. We propose a new image-level scoring process to rank unlabeled images for their automatic selection, which clearly outperforms classical scores. The proposed method can be applied to videos and sets of still images. In the former case, temporal selection rules can complement our scoring process. As a relevant use case, we extensively study the performance of our method on the task of pedestrian detection. Overall, the experiments show that the proposed method performs better than random selection.

Automatic classification of foods is a way to control food intake and tackle with obesity. However, it is a challenging problem since foods are highly deformable and complex objects. Results on ImageNet dataset have revealed that Convolutional Neural Network has a great expressive power to model natural objects. Nonetheless, it is not trivial to train a ConvNet from scratch for classification of foods. This is due to the fact that ConvNets require large datasets and to our knowledge there is not a large public dataset of food for this purpose. Alternative solution is to transfer knowledge from trained ConvNets to the domain of foods. In this work, we study how transferable are state-of-art ConvNets to the task of food classification. We also propose a method for transferring knowledge from a bigger ConvNet to a smaller ConvNet by keeping its accuracy similar to the bigger ConvNet. Our experiments on UECFood256 datasets show that Googlenet, VGG and residual networks produce comparable results if we start transferring knowledge from appropriate layer. In addition, we show that our method is able to effectively transfer knowledge to the smaller ConvNet using unlabeled samples.

Convolutional Neural Networks (CNN) beat the human performance on German Traffic Sign Benchmark competition. Both the winner and the runner-up teams trained CNNs to recognize 43 traffic signs. However, both networks are not computationally efficient since they have many free parameters and they use highly computational activation functions. In this paper, we propose a new architecture that reduces the number of the parameters 27% and 22% compared with the two networks. Furthermore, our network uses Leaky Rectified Linear Units (ReLU) as the activation function that only needs a few operations to produce the result. Specifically, compared with the hyperbolic tangent and rectified sigmoid activation functions utilized in the two networks, Leaky ReLU needs only one multiplication operation which makes it computationally much more efficient than the two other functions. Our experiments on the Gertman Traffic Sign Benchmark dataset shows 0:6% improvement on the best reported classification accuracy while it reduces the overall number of parameters 85% compared with the winner network in the competition.

People are getting more interested to record their sport activities using head-worn or hand-held cameras. This type of videos which is called egocentric sport videos has different motion and appearance patterns compared with life-logging videos. While a life-logging video can be defined in terms of well-defined human-object interactions, notwithstanding, it is not trivial to describe egocentric sport videos using well-defined activities. For this reason, summarizing egocentric sport videos based on human-object interaction might fail to produce meaningful results. In this paper, we propose an unsupervised method for summarizing egocentric videos by identifying the key-frames of the video. Our method utilizes both appearance and motion information and it automatically finds the number of the key-frames. Our blind user study on the new dataset collected from YouTube shows that in 93:5% cases, the users choose the proposed method as their first video summary choice. In addition, our method is within the top 2 choices of the users in 99% of studies.

Controlling the food intake is an efficient way that each person can undertake to tackle the obesity problem in countries worldwide. This is achievable by developing a smartphone application that is able to recognize foods and compute their calories. State-of-art methods are chiefly based on hand-crafted feature extraction methods such as HOG and Gabor. Recent advances in large-scale object recognition datasets such as ImageNet have revealed that deep Convolutional Neural Networks (CNN) possess more representation power than the hand-crafted features. The main challenge with CNNs is to find the appropriate architecture for each problem. In this paper, we propose a deep CNN which consists of 769; 988 parameters. Our experiments show that the proposed CNN outperforms the state-of-art methods and improves the best result of traditional methods 17%. Moreover, using an ensemble of two CNNs that have been trained two different times, we are able to improve the classification performance 21:5%.

Segmentation of the breast region is usually the first step in the analysis of mammograms. Due to the non-uniformity of the background, breast segmentation presents several difficulties especially for film based mammograms. Our experimental results show that 50% of digitized film based mammograms in the mini-MIAS database do not have uniform intensity in the background. For this reason, applying a global thresholding method produces inaccurate results. In addition, finding the optimal global threshold value by only using histogram information requires a reliable objective function that characterizes the statistics of the background and the mammogram regions in the digitized mammograms. A second way to find the boundary of the breast consists in fitting a deformable model, such as snakes, on the mammogram. However, this method has three main shortcomings. First, the model must be initialized near the boundary. Second, using gradient information in the objective function can push the boundary toward the tissues inside the breast rather than the actual boundary. Third, in some mammograms the breast region is occluded by artifacts, such as labels, that have high gradient values on their boundary and cause the deformable model to be fitted on the artifact. To address these problems we propose a probabilistic adaptive thresholding method that uses texture information and its probability to find the most probable threshold values for specific parts of the mammogram. The experimental results on mini-MIAS database show that our proposed method outperforms the state-of-art methods and improves the accuracy at least 37% in comparison with the best results obtained by contour growing methods.