Abstract: Avoiding vehicle-to-pedestrian crashes is a critical requirement for nowadays advanced driver assistant systems (ADAS) and future self-driving vehicles. Accordingly, detecting pedestrians from raw sensor data has a history of more than 15 years of research, with vision playing a central role.
During the last years, deep learning has boosted the accuracy of image-based pedestrian detectors.
However, detection is just the first step towards answering the core question, namely is the vehicle going to crash with a pedestrian provided preventive actions are not taken? Therefore, knowing as soon as possible if a detected pedestrian has the intention of crossing the road ahead of the vehicle is
essential for performing safe and comfortable maneuvers that prevent a crash. However, compared to pedestrian detection, there is relatively little literature on detecting pedestrian intentions. This paper aims to contribute along this line by presenting a new vision-based approach which analyzes the
pose of a pedestrian along several frames to determine if he or she is going to enter the road or not. We present experiments showing 750 ms of anticipation for pedestrians crossing the road, which at a typical urban driving speed of 50 km/h can provide 15 additional meters (compared to a pure pedestrian detector) for vehicle automatic reactions or to warn the driver. Moreover, in contrast with state-of-the-art methods, our approach is monocular, neither requiring stereo nor optical flow information.

Abstract: Before the big data era, scene recognition was often approached with two-step inference using localized intermediate representations (objects, topics, and so on). One of such approaches is the semantic manifold (SM), in which patches and images are modeled as points in a semantic probability simplex. Patch models are learned resorting to weak supervision via image labels, which leads to the problem of scene categories co-occurring in this semantic space. Fortunately, each category has its own co-occurrence patterns that are consistent across the images in that category. Thus, discovering and modeling these patterns are critical to improve the recognition performance in this representation. Since the emergence of large data sets, such as ImageNet and Places, these approaches have been relegated in favor of the much more powerful convolutional neural networks (CNNs), which can automatically learn multi-layered representations from the data. In this paper, we address many limitations of the original SM approach and related works. We propose discriminative patch representations using neural networks and further propose a hybrid architecture in which the semantic manifold is built on top of multiscale CNNs. Both representations can be computed significantly faster than the Gaussian mixture models of the original SM. To combine multiple scales, spatial relations, and multiple features, we formulate rich context models using Markov random fields. To solve the optimization problem, we analyze global and local approaches, where a top-down hierarchical algorithm has the best performance. Experimental results show that exploiting different types of contextual relations jointly consistently improves the recognition accuracy.

Abstract: Scene recognition with RGB images has been extensively studied and has reached very remarkable recognition levels, thanks to convolutional neural networks (CNN) and large scene datasets. In contrast, current RGB-D scene data is much more limited, so often leverages RGB large datasets, by transferring pretrained RGB CNN models and fine-tuning with the target RGB-D dataset. However, we show that this approach has the limitation of hardly reaching bottom layers, which is key to learn modality-specific features. In contrast, we focus on the bottom layers, and propose an alternative strategy to learn depth features combining local weakly supervised training from patches followed by global fine tuning with images. This strategy is capable of learning very discriminative depth-specific features with limited depth images, without resorting to Places-CNN. In addition we propose a modified CNN architecture to further match the complexity of the model and the amount of data available. For RGB-D scene recognition, depth and RGB features are combined by projecting them in a common space and further leaning a multilayer classifier, which is jointly optimized in an end-to-end network. Our framework achieves state-of-the-art accuracy on NYU2 and SUN RGB-D in both depth only and combined RGB-D data.

Abstract: Multispectral images captured with a single sensor camera have become an attractive alternative for numerous computer vision applications. However, in order to fully exploit their potentials, the color restoration problem (RGB representation) should be addressed. This problem is more evident in outdoor scenarios containing vegetation, living beings, or specular materials. The problem of color distortion emerges from the sensitivity of sensors due to the overlap of visible and near infrared spectral bands. This paper empirically evaluates the variability of the near infrared (NIR) information with respect to the changes of light throughout the day. A tiny neural network is proposed to restore the RGB color representation from the given RGBN (Red, Green, Blue, NIR) images. In order to evaluate the proposed algorithm, different experiments on a RGBN outdoor dataset are conducted, which include various challenging cases. The obtained result shows the challenge and the importance of addressing color restoration in single sensor multispectral images.

Abstract: This paper considers the problem of recipe-oriented image-ingredient correlation learning with multi-attributes for recipe retrieval and exploration. Existing methods mainly focus on food visual information for recognition while we model visual information, textual content (e.g., ingredients), and attributes (e.g., cuisine and course) together to solve extended recipe-oriented problems, such as multimodal cuisine classification and attribute-enhanced food image retrieval. As a solution, we propose a multimodal multitask deep belief network (M3TDBN) to learn joint image-ingredient representation regularized by different attributes. By grouping ingredients into visible ingredients (which are visible in the food image, e.g., “chicken” and “mushroom”) and nonvisible ingredients (e.g., “salt” and “oil”), M3TDBN is capable of learning both midlevel visual representation between images and visible ingredients and nonvisual representation. Furthermore, in order to utilize different attributes to improve the intermodality correlation, M3TDBN incorporates multitask learning to make different attributes collaborate each other. Based on the proposed M3TDBN, we exploit the derived deep features and the discovered correlations for three extended novel applications: 1) multimodal cuisine classification; 2) attribute-augmented cross-modal recipe image retrieval; and 3) ingredient and attribute inference from food images. The proposed approach is evaluated on the constructed Yummly dataset and the evaluation results have validated the effectiveness of the proposed approach.

Abstract: Historical records of daily activities provide intriguing insights into the life of our ancestors, useful for demographic and genealogical research. For example, marriage license books have been used for centuries by ecclesiastical and secular institutions to register marriages. These books follow a simple structure of the text in the records with a evolutionary vocabulary, mainly composed of proper names that change along the time. This distinct vocabulary makes automatic transcription and semantic information extraction difficult tasks. In previous works we studied the use of category-based language models and how a Grammatical Inference technique known as MGGI could improve the accuracy of these tasks. In this work we analyze the main causes of the semantic errors observed in previous results and apply a better implementation of the MGGI technique to solve these problems. Using the resulting language model, transcription and information extraction experiments have been carried out, and the results support our proposed approach.

Abstract: arXiv:1707.01294
This paper proposes a region based convolutional neural network for segmentation-free word spotting. Our network takes as input an image and a set of word candidate bound- ing boxes and embeds all bounding boxes into an embedding space, where word spotting can be casted as a simple nearest neighbour search between the query representation and each of the candidate bounding boxes. We make use of PHOC embedding as it has previously achieved significant success in segmentation- based word spotting. Word candidates are generated using a simple procedure based on grouping connected components using some spatial constraints. Experiments show that R-PHOC which operates on images directly can improve the current state-of- the-art in the standard GW dataset and performs as good as PHOCNET in some cases designed for segmentation based word spotting.

Abstract: Developing a text detection method which is invariant to scripts in natural scene images is a challeng- ing task due to different geometrical structures of various scripts. Besides, multi-oriented of text lines in natural scene images make the problem more challenging. This paper proposes to explore ring radius transform (RRT) for text detection in multi-oriented and multi-script environments. The method finds component regions based on convex hull to generate radius matrices using RRT. It is a fact that RRT pro- vides low radius values for the pixels that are near to edges, constant radius values for the pixels that represent stroke width, and high radius values that represent holes created in background and convex hull because of the regular structures of text components. We apply k -means clustering on the radius matrices to group such spatially coherent regions into individual clusters. Then the proposed method studies the radius values of such cluster components that are close to the centroid and far from the cen- troid to detect text components. Furthermore, we have developed a Bangla dataset (named as ISI-UM dataset) and propose a semi-automatic system for generating its ground truth for text detection of arbi- trary orientations, which can be used by the researchers for text detection and recognition in the future. The ground truth will be released to public. Experimental results on our ISI-UM data and other standard datasets, namely, ICDAR 2013 scene, SVT and MSRA data, show that the proposed method outperforms the existing methods in terms of multi-lingual and multi-oriented text detection ability.