Abstract

In this paper, we improve the performance of the recently proposed Direct Query Classifier (DQC). The (DQC) is a classifier based retrieval method and in general, such methods have been shown to be superior to the OCR-based solutions for performing retrieval in many practical document image datasets. In (DQC), the classifiers are trained for a set of frequent queries and seamlessly extended for the rare and arbitrary queries. This extends the classifier based retrieval paradigm to an unlimited number of classes (words) present in a language. The (DQC) requires indexing cut-portions (n-grams) of the word image and DTW distance has been used for indexing. However, DTW is computationally slow and therefore limits the performance of the (DQC). We introduce query specific DTW distance, which enables effective computation of global principal alignments for novel queries. Since the proposed query specific DTW distance is a linear approximation of the DTW distance, it enhances the performance of the (DQC). Unlike previous approaches, the proposed query specific DTW distance uses both the class mean vectors and the query information for computing the global principal alignments for the query. Since the proposed method computes the global principal alignments using n-grams, it works well for both frequent and rare queries. We also use query expansion (QE) to further improve the performance of our query specific DTW. This also allows us to seamlessly adapt our solution to new fonts, styles and collections. We have demonstrated the utility of the proposed technique over 3 different datasets. The proposed query specific DTW performs well compared to the previous DTW approximations.

Abstract

We present an approach for unsupervised training of CNNs in order to learn discriminative face representations. We mine supervised training data by noting that multiple faces in the same video frame must belong to different persons and the same face tracked across multiple frames must belong to the same person. We obtain millions of face pairs from hundreds of videos without using any manual supervision. Although faces extracted from videos have a lower spatial resolution than those which are available as part of standard supervised face datasets such as LFW and CASIA-WebFace, the former represent a much more realistic setting, e.g. in surveillance scenarios where most of the faces detected are very small. We train our CNNs with the relatively low resolution faces extracted from video frames collected, and achieve a higher verification accuracy on the benchmark LFW dataset cf. hand-crafted features such as LBPs, and even surpasses the performance of state-of-the-art deep networks such as VGG-Face, when they are made to work with low resolution input images.

Abstract

The immense success of deep learning based methods in computer vision heavily relies on large scale training datasets. These richly annotated datasets help the network learn discriminative visual features. Collecting and annotating such datasets requires a tremendous amount of human effort and annotations are limited to popular set of classes. As an alternative, learning visual features by designing auxiliary tasks which make use of freely available self-supervision has become increasingly popular in the computer vision community. In this paper, we put forward an idea to take advantage of multi-modal context to provide self-supervision for the training of computer vision algorithms. We show that adequate visual features can be learned efficiently by training a CNN to predict the semantic textual context in which a particular image is more probable to appear as an illustration. More specifically we use popular text embedding techniques to provide the self-supervision for the training of deep CNN.

Abstract

We consider the task of learning to estimate human pose in still images. In order to avoid the high cost of full supervision, we propose to use a diverse data set, which consists of two types of annotations: (i) a small number of images are labeled using the expensive ground-truth pose; and (ii) other images are labeled using the inexpensive action label. As action information helps narrow down the pose of a human, we argue that this approach can help reduce the cost of training without significantly affecting the accuracy. To demonstrate this we design a probabilistic framework that employs two distributions: (i) a conditional distribution to model the uncertainty over the human pose given the image and the action; and (ii) a prediction distribution, which provides the pose of an image without using any action information. We jointly estimate the parameters of the two aforementioned distributions by minimizing their dissimilarity coefficient, as measured by a task-specific loss function. During both training and testing, we only require an efficient sampling strategy for both the aforementioned distributions. This allows us to use deep probabilistic networks that are capable of providing accurate pose estimates for previously unseen images. Using the MPII data set, we show that our approach outperforms baseline methods that either do not use the diverse annotations or rely on pointwise estimates of the pose.

Abstract

We are motivated by the fact that multiple representations of the environment are required to stand for the changes in appearance with time and for changes that appear in a cyclic manner. These changes are, for example, from day to night time, and from day to day across seasons. In such situations, the robot visits the same routes multiple times and collects different appearances of it. Multiple visual experiences usually find robotic vision applications like visual localization, mapping, place recognition, and autonomous navigation. The novelty in this paper is an algorithm that connects multiple visual experiences via aligning multiple image sequences. This problem is solved by finding the maximum flow in a directed graph flow-network, whose vertices represent the matches between frames in the test and reference sequences. Edges of the graph represent the cost of these matches. The problem of finding the best match is reduced to finding the minimum-cut surface, which is solved as a maximum flow network problem. Application to visual localization is considered in this paper to show the effectiveness of the proposed multiple image sequence alignment method, without loosing its generality. Experimental evaluations show that the precision of sequence matching is improved by considering multiple visual sequences for the same route, and that the method performs favorably against state-of-the-art single representation methods like SeqSLAM and ABLE-M.

Abstract

t—Deep neural networks for image classification typically consists of a convolutional feature extractor followed by a fully connected classifier network. The predicted and the ground truth labels are represented as one hot vectors. Such a representation assumes that all classes are equally dissimilar. However, classes have visual similarities and often form a hierarchy. Learning this latent hierarchy explicitly in the architecture could provide invaluable insights. We propose an alternate architecture to the classifier network called the Latent Hierarchy (LH) Classifier and an end to end learned Class2Str mapping which discovers a latent hierarchy of the classes. We show that for some of the best performing architectures on CIFAR and Imagenet datasets, the proposed replacement and training by LH classifier recovers the accuracy, with a fraction of the number of parameters in the classifier part. Compared to the previous work of HDCNN, which also learns a 2 level hierarchy, we are able to learn a hierarchy at an arbitrary number of levels as well as obtain an accuracy improvement on the Imagenet classification task over them. We also verify that many visually similar classes are grouped together, under the learnt hierarchy

Abstract

Abstract—Convolutional Neural Networks have been shown to be promising for image tampering detection in the recent years. However, the number of tampered images available to train a network is still small. This is mainly due to the cumbersomeness involved in creating lots of tampered images. As a result, the potential offered by these networks is not completely exploited. In this work, we propose a simple method to address this problem by augmenting data using inpainting and compositing schemes. We consider different forms of inpainting like simple inpainting and semantic inpainting as well as compositing schemes like feathering in order to augment the data. A domain adaptation technique is employed to reduce the domain shift between the augmented data and the data available using proprietary softwares. We demonstrate that this method of augmentation is effective in improving the detection accuracies. We present experimental evaluation on two popular datasets for image tampering detection to demonstrate the effectiveness of the proposed approach.

Abstract

With recent advances in the field of computer vision and especially deep learning, many fully connected and convolutional neural networks have been trained to achieve state-of-the-art performance on a wide variety of tasks such as speech recognition, image classification and natural language processing. For classification tasks however, most of these deep learning models employ the softmax activation function for prediction and minimize cross-entropy loss. In contrast, we demonstrate a consistent advantage by replacing the softmax layer by a set of binary SVM classifiers organized in a tree or DAG (Directed Acyclic Graph) structure. The idea is to not treat the multiclass classification problem as a whole but to break it down into smaller binary problems where each classifier acts as an expert by focusing on differentiating between only two classes and thus improves the overall accuracy. Furthermore, by arranging the classifiers in a DAG structure, we later also show how it is possible to further improve the performance of the binary classifiers by learning more discriminative features through the same deep network. We validated the proposed methodology on two benchmark datasets, and the results corroborated our claim.

Abstract

Overhead imageries play a crucial role in many applications such as urban planning, crop yield forecasting, mapping, and policy making. Semantic segmentation could enable automatic, efficient, and large-scale understanding of overhead imageries for these applications. However, semantic segmentation of overhead imageries is a challenging task, primarily due to the large domain gap from existing research in ground imageries, unavailability of large-scale dataset with pixel-level annotations, and inherent complexity in the task. Readily available vast amount of unlabeled overhead imageries share more common structures and patterns compared to the ground imageries, therefore, its large-scale analysis could benefit from unsupervised feature learning techniques. In this work, we study various self-supervised feature learning techniques for semantic segmentation of overhead imageries. We choose image semantic inpainting as a self-supervised task [36] for our experiments due to its proximity to the semantic segmentation task. We (i) show that existing approaches are inefficient for semantic segmentation, (ii) propose architectural changes towards self-supervised learning for semantic segmentation, (iii) propose an adversarial training scheme for self-supervised learning by increasing the pretext task’s difficulty gradually and show that it leads to learning better features, and (iv) propose a unified approach for overhead scene parsing, road network extraction, and land cover estimation. Our approach improves over training from scratch by more than 10% and ImageNet pre-trained network by more than 5% mIOU.

Abstract

Semantic segmentation is a popular task in computer vision today, and deep neural network models have emerged as the popular solution to this problem in recent times. The typical loss function used to train neural networks for this task is the cross-entropy loss. However, the success of the learned models is measured using Intersection-OverUnion (IoU), which is inherently non-differentiable. This gap between performance measure and loss function results in a fall in performance, which has also been studied by few recent efforts. In this work, we propose a novel method to automatically learn a surrogate loss function that approximates the IoU loss and is better suited for good IoU performance. To the best of our knowledge, this is the first such work that attempts to learn a loss function for this purpose. The proposed loss can be directly applied over any network. We validated our method over different networks (FCN, SegNet, UNet) on the PASCAL VOC and Cityscapes datasets. Our results on this work show consistent improvement over baseline methods.