Abstract

Most of the Indian scripts do not have any robust commercial OCRs. Many of the laboratory prototypes report reasonable results at recognition/classification stage. However, word level accuracies are still poor. It is well known that word accuracy decreases as the number of characters in a word increase. For Malayalam, the average number of characters in a word is almost twice that of English. Moreover, the number of words required to cover 80% of the Malayalam language is more than forty times that of other Indian languages such as Hindi. Hence a direct dictionary based post-processing scheme is not suitable for Malayalam. In this paper, we propose a post-processing scheme which uses statistical language models at the sub-character level to boost word level recognition results. We use a multi-stage graph representation and formulate the recognition task as an optimization problem. Edges of the graph encode the language information and nodes represent the visual similarities. An optimal path from source node to destination node represents the recognized text. We validate our method on more than 10,000 words from a Malayalam corpus.

Abstract

Matching word images has many applications in document recognition and retrieval systems. Dynamic Time Warping (DTW) is popularly used to estimate the similarity between word images. Word images are represented as sequences of feature vectors, and the cost associated with dynamic programming based alignment is considered as the dissimilarity between them. However, such approaches are computationally costly when compared to fixed length matching schemes. In this paper, we explore systematic methods for identifying appropriate distance metrics for a given database or language. This is achieved by learning query specific distance functions which can be computed online efficiently. We show that a weighted Euclidean distance can outperform DTW for matching word images. This class of distance functions are also ideal for scalability and large scale matching. Our results are validated with mean Average Precision (mAP) on a fully annotated data set of 160K word images. We then show that the learnt distance functions can even be extended to a new database to obtain accurate retrieval.

Abstract

Popular image retrieval schemes generally rely only on a single mode, (either low level visual features or embedded text) for searching in multimedia databases. Many popular image collections (eg. those emerging over Internet) have associated tags, often for human consumption. A natural extension is to combine information from multiple modes for enhancing effectiveness in retrieval. In this paper, we propose two techniques: Multi-modal Latent Semantic Indexing (MMLSI) and Multi-Modal Probabilistic Latent Semantic Analysis (MMpLSA). These methods are obtained by directly extending their traditional single mode counter parts. Both these methods incorporate visual features and tags by generating simultaneous semantic contexts. The experimental results demonstrate an improved accuracy over other single and multi-modal methods

Abstract

Kernel methods yield state-of-the-art performance in certain applications such as image classification and object detection. However, large scale problems require machine learning techniques of at most linear complexity and these are usually limited to linear kernels. This unfortunately rules out gold-standard kernels such as the generalized RBF kernels (e.g. exponential-χ 2 ). Recently, Maji and Berg [13] and Vedaldi and Zisserman [20] proposed explicit feature maps to approximate the additive kernels (intersection, χ 2 , etc.) by linear ones, thus enabling the use of fast machine learning technique in a non-linear context. An analogous technique was proposed by Rahimi and Recht [14] for the translation invariant RBF kernels. In this paper, we complete the construction and combine the two techniques to obtain explicit feature maps for the generalized RBF kernels. Furthermore, we investigate a learning method using l 1 regularization to encourage sparsity in the final vector representation, and thus reduce its dimension. We evaluate this technique on the VOC 2007 detection challenge, showing when it can improve on fast additive kernels, and the trade-offs in complexity and accuracy

Abstract

In this paper, we present a monocular camera based terrain classification scheme. The uniqueness of the proposed scheme is that it inherently incorporates spatial smoothness while segmenting a image, without requirement of post-processing smoothing methods. The algorithm is extremely fast because it is build on top of a Random Forest classifier. The baseline algorithm uses color, texture and their combination with classifiers such as SVM and Random Forests. We present comparison across features and classifiers. We further enhance the algorithm through a label transfer method. The efficacy of the proposed solution can be seen as we reach a low error rates on both our dataset and other publicly available datasets.

Abstract

Semantic analysis of a document collection can be viewed as an unsupervised clustering of the constituent words and documents around hidden or latent concepts. This has shown to improve the performance of visual bag of words in image retrieval. However, the enhancement in performance depends heavily on the right choice of number of semantic concepts. Most of the semantic indexing schemes are also computationally costly. In this paper, we employ a bipartite graph model (BGM) for image retrieval. BGM is a scalable datastructure that aids semantic indexing in an efficient manner. It can also be incrementally updated. BGM uses tf-idf values for building a semantic bipartite graph. We also introduce a graph partitioning algorithm that works on the BGM to retrieve semantically relevant images from a database. We demonstrate the properties as well as performance of our semantic indexing scheme through a series of experiments. We also compare our methods with incremental pLSA

Abstract

An important pre-requisite for many tasks like Visual Servoing and visual SLAM is the task of tracking the underlying features. The use of planar features for these purposes has gained importance recently. Complementing current planar tracking works in the robotics literature, which use multiple features, we formulate the tracking problem using multiple planes. Inspired by the maturity in understanding of geometric quantities like the homography in computer vision, we develop a system based on the Unscented Kalman Filter (UKF) that localizes the camera and estimates the plane parameters of a scene, using homographies as measurement. Homographies are estimated using tracked feature points. We show that this framework provides significant robustness and stability to the system under significant changes of illumination, occlusion etc. Finally, we also propose a Convex optimization based solution for the initialization of this system, which is capable of producing globally optimal estimates, and is a useful algorithm in its own right. Several synthetic and real results are presented to demonstrate the efficacy of our approach.

Abstract

In this paper, we present a practical vision based Simultaneous Localization and Mapping (SLAM) system for a highly dynamic environment. We adopt a multibody Structure from Motion (SfM) approach, which is the generalization of classical SfM to dynamic scenes with multiple rigidly moving objects. The proposed framework of multibody visual SLAM allows choosing between full 3D reconstruction or simply tracking of the moving objects, which adds flexibility to the system, for scenes containing non-rigid objects or objects having insufficient features for reconstruction. The solution demands a motion segmentation framework that can segment feature points belonging to different motions and maintain the segmentation with time. We propose a realtime incremental motion segmentation algorithm for this purpose. The motion segmentation is robust and is capable of segmenting difficult degenerate motions, where the moving objects is followed by a moving camera in the same direction. This robustness is attributed to the use of efficient geometric constraints and a probability framework which propagates the uncertainty in the system. The motion segmentation module is tightly coupled with feature tracking and visual SLAM, by exploring various feed-backs in between these modules. The integrated system can simultaneously perform realtime visual SLAM and tracking of multiple moving objects using only a single monocular camera.

Abstract

In this paper, we present an approach to discover characteristic patterns in videos. We characterize the videos based on frequently occurring patterns like scenes, characters, sequence of frames in an unsupervised setting. With our approach, we are able to detect the representative scenes and characters of movies. We also present a method for detecting video stop-words in broadcast news videos based on the frequency of occurrence of sequence of frames. These are analogous to stop-words in text classification and search. We employ two different video mining schemes; both aimed at detecting frequent and representative patterns. For one of our mining approaches, we use an efficient frequent pattern mining algorithm over a quantized feature space. Our second approach uses a Random Forest to first represent video data as sequences, and then mine the frequent patterns. We validate the proposed approaches on broadcast news videos and our database of 81 Oscar winning movies.

Abstract

One of the most common computer vision tasks is that of recognizing the category of objects present in a given image. Previous work has mostly focused on building accurate classifiers based on carefully selected features. Classification is often carried on individual test images, while most of the practical situations, such as webscale image indexing, demand the simultaneous classification of a large collection of images. This is especially true for real-world datasets, that already contain numerous un-indexed images and videos. In this paper, we work towards developing a computationally efficient approach towards object recognition, that is inspired by retrieval schemes. We perform an offline indexing of the features from the collection, so that the classifier only needs to work on a small subset of the entire feature set. Over a set of 2 Million features extracted from 7000 images, classification against 5 object categories using a standard SVM would require more than 260 hours. Over the same test case, the classification time using our indexing based approach is reduced to less than 13 hours. The compromise on the accuracy is less than 7% for the 20X speedup achieved.