Abstract: Food image recognition has recently attracted the attention of many researchers, due to the challenging problem it poses, the ease collection of food images, and its numerous applications to health and leisure. In real applications, it is necessary to analyze and recognize thousands of different foods. For this purpose, we propose a novel prediction scheme based on a class hierarchy that considers local classifiers, in addition to a flat classifier. In order to make a decision about which approach to use, we define different criteria that take into account both the analysis of the Epistemic Uncertainty estimated from the ‘children’ classifiers and the prediction from the ‘parent’ classifier. We evaluate our proposal using three Uncertainty estimation methods, tested on two public food datasets. The results show that the proposed method reduces parent-child error propagation in hierarchical schemes and improves classification results compared to the single flat classifier, meanwhile maintains good performance regardless the Uncertainty estimation method chosen.

Abstract: This paper proposes a technique for RGB-D scene segmentation using Multi-class
Multi-scale Stacked Sequential Learning (MMSSL) paradigm. Following recent trends in state-of-the-art, a base classifier uses an initial SLIC segmentation to obtain superpixels which provide a diminution of data while retaining object boundaries. A series of color and depth features are extracted from the superpixels, and are used in a Conditional Random Field (CRF) to predict superpixel labels. Furthermore, a Random Forest (RF) classifier using random offset features is also used as an input to the CRF, acting as an initial prediction. As a stacked classifier, another Random Forest is used acting on a spatial multi-scale decomposition of the CRF confidence map to correct the erroneous labels assigned by the previous classifier. The model is tested on the popular NYU-v2 dataset.
The approach shows that simple multi-modal features with the power of the MMSSL
paradigm can achieve better performance than state of the art results on the same dataset.

Abstract: Graph-based representations are experiencing a growing usage in visual recognition and retrieval due to their representational power in front of classical appearance-based representations. However, retrieving a query graph from a large dataset of graphs implies a high computational complexity. The most important property for a large-scale retrieval is the search time complexity to be sub-linear in the number of database examples. With this aim, in this paper we propose a graph indexation formalism applied to visual retrieval. A binary embedding is defined as hashing keys for graph nodes. Given a database of labeled graphs, graph nodes are complemented with vectors of attributes representing their local context. Then, each attribute vector is converted to a binary code applying a binary-valued hash function. Therefore, graph retrieval is formulated in terms of finding target graphs in the database whose nodes have a small Hamming distance from the query nodes, easily computed with bitwise logical operators. As an application example, we validate the performance of the proposed methods in different real scenarios such as handwritten word spotting in images of historical documents or symbol spotting in architectural floor plans.

Abstract: We present the results of recent challenges in Automated Computer Vision (AutoCV, renamed here for clarity AutoCV1 and AutoCV2, 2019), which are part of a series of challenge on Automated Deep Learning (AutoDL). These two competitions aim at searching for fully automated solutions for classification tasks in computer vision, with an emphasis on any-time performance. The first competition was limited to image classification while the second one included both images and videos. Our design imposed to the participants to submit their code on a challenge platform for blind testing on five datasets, both for training and testing, without any human intervention whatsoever. Winning solutions adopted deep learning techniques based on already published architectures, such as AutoAugment, MobileNet and ResNet, to reach state-of-the-art performance in the time budget of the challenge (only 20 minutes of GPU time). The novel contributions include strategies to deliver good preliminary results at any time during the learning process, such that a method can be stopped early and still deliver good performance. This feature is key for the adoption of such techniques by data analysts desiring to obtain rapidly preliminary results on large datasets and to speed up the development process. The soundness of our design was verified in several aspects: (1) Little overfitting of the on-line leaderboard providing feedback on 5 development datasets was observed, compared to the final blind testing on the 5 (separate) final test datasets, suggesting that winning solutions might generalize to other computer vision classification tasks; (2) Error bars on the winners’ performance allow us to say with confident that they performed significantly better than the baseline solutions we provided; (3) The ranking of participants according to the any-time metric we designed, namely the Area under the Learning Curve, was different from that of the fixed-time metric, i.e. AUC at the end of the fixed time budget. We released all winning solutions under open-source licenses. At the end of the AutoDL challenge series, all data of the challenge will be made publicly available, thus providing a collection of uniformly formatted datasets, which can serve to conduct further research, particularly on meta-learning.

Abstract: Images with visual and scene text content are ubiquitous in everyday life. However, current image interpretation systems are mostly limited to using only the visual features, neglecting to leverage the scene text content. In this paper, we propose to jointly use scene text and visual channels for robust semantic interpretation of images. We do not only extract and encode visual and scene text cues, but also model their interplay to generate a contextual joint embedding with richer semantics. The contextual embedding thus generated is applied to retrieval and classification tasks on multimedia images, with scene text content, to demonstrate its effectiveness. In the retrieval framework, we augment our learned text-visual semantic representation with scene text cues, to mitigate vocabulary misses that may have occurred during the semantic embedding. To deal with irrelevant or erroneous recognition of scene text, we also apply query-based attention to our text channel. We show how the multi-channel approach, involving visual semantics and scene text, improves upon state of the art.

Abstract: Online continual learning aims to learn from a non-IID stream of data from a number of different tasks, where the learner is only allowed to consider data once. Methods are typically allowed to use a limited buffer to store some of the images in the stream. Recently, it was found that feature replay, where an intermediate layer representation of the image is stored (or generated) leads to superior results than image replay, while requiring less memory. Quantized exemplars can further reduce the memory usage. However, a drawback of these methods is that they use a fixed (or very intransigent) backbone network. This significantly limits the learning of representations that can discriminate between all tasks. To address this problem, we propose an auxiliary classifier auto-encoder (ACAE) module for feature replay at intermediate layers with high compression rates. The reduced memory footprint per image allows us to save more exemplars for replay. In our experiments, we conduct task-agnostic evaluation under online continual learning setting and get state-of-the-art performance on ImageNet-Subset, CIFAR100 and CIFAR10 dataset.

Abstract: During the last years, graph-based representations are experiencing a growing usage in visual recognition and retrieval due to their ability to capture both structural and appearance-based information. Thus, they provide a greater representational power than classical statistical frameworks. However, graph-based representations leads to high computational complexities usually dealt by graph embeddings or approximated matching techniques. Despite their representational power, they are very sensitive to noise and small variations of the input image. With the aim to cope with the time complexity and the variability present in the generated graphs, in this paper we propose to construct a novel hierarchical graph representation. Graph clustering techniques adapted from social media analysis have been used in order to contract a graph at different abstraction levels while keeping information about the topology. Abstract nodes attributes summarise information about the contracted graph partition. For the proposed representations, a coarse-to-fine matching technique is defined. Hence, small graphs are used as a filtering before more accurate matching methods are applied. This approach has been validated in real scenarios such as classification of colour images or retrieval of handwritten words (i.e. word spotting).

Abstract: Class-specific text proposal algorithms can efficiently reduce the search space for possible text object locations in an image. In this paper we combine the Text Proposals algorithm with Fully Convolutional Networks to efficiently reduce the number of proposals while maintaining the same recall level and thus gaining a significant speed up. Our experiments demonstrate that such text proposal approaches yield significantly higher recall rates than state-of-the-art text localization techniques, while also producing better-quality localizations. Our results on the ICDAR 2015 Robust Reading Competition (Challenge 4) and the COCO-text datasets show that, when combined with strong word classifiers, this recall margin leads to state-of-the-art results in end-to-end scene text recognition.