Sanskrit is a highly composite language, morphologically and phonetically complex. One of the major challenges in processing Sanskrit is the splitting of compound words that are merged phonetically. Recognizing the exact location of splits in a compound word is difficult since several possible splits can be found, but only a few of them are semantically meaningful. This paper proposes a novel deep learning method that uses two bi-encoders and a multi-head attention module to predict the valid split location in Sanskrit compound words. The two bi-encoders process the input sequence in direct and reverse order respectively. The model learns the character-level context in which the splitting occurs by exploiting the correlation between the direct and reverse dynamics of the characters sequence. The results of the proposed model are compared with a stateof-the-art technique that adopts a bidirectional recurrent network to solve the same task. Experimental results show that the proposed model correctly identifies where the compound word should be split into its components in 89.27% of cases, outperforming the state-of-the-art technique. The paper also proposes a dataset developed from the repository of the Digital Corpus of Sanskrit (DCS) and the University of Hyderabad (UoH) corpus.

Abstract Predicting pedestrian crossing intentions is crucial for enhancing autonomous vehicles’ safety and decision-making capabilities. Timely and accurate predictions made well before potential crossing events are essential for enabling vehicles to respond appropriately and prevent accidents. Unlike state-of-the-art models that rely on complex architectures and multiple input modalities, this article proposes Context-Ped, a minimalist effective model for pedestrian intention prediction. To capture critical contextual information, Context-Ped processes two key visual inputs: cropped pedestrian images and environmental road structure. By integrating residual networks with recurrent architectures, the model extracts robust spatiotemporal features through the aggregation of hidden states from ConvLSTM layers. This operation serves as a temporal integration strategy that accumulates the spatiotemporal patterns over the sequence, capturing both dynamic motion cues and static contextual information to encode both pedestrian behavior and environmental context. As a result, Context-Ped achieves accurate intention prediction while maintaining computational efficiency. Additionally, the proposed approach addresses the often-overlooked issue of class imbalance by employing the focal loss function, which significantly enhances performance for the minority class. For evaluation, AUC-ROC and recall metrics are prioritized over traditional accuracy, providing clearer insight into the model’s performance. Experimental results on the JAAD and PIE datasets demonstrate competitive performance, achieving an AUC of 79% and a crossing recall of 82% on the JAAD dataset, outperforming more complex state-of-the-art models and by only using visual input data.

Abstract This paper proposes a new method to improve the training efficiency of deep convolutional neural networks. During training, the method evaluates scores to measure how much each layer’s parameters change and whether the layer will continue learning or not. Based on these scores, the network is scaled down such that the number of parameters to be learned is reduced, yielding a speed-up in training. Unlike state-of-the-art methods that try to compress the network to be used in the inference phase or to limit the number of operations performed in the back-propagation phase, the proposed method is novel in that it focuses on reducing the number of operations performed by the network in the forward propagation during training. The proposed training strategy has been validated on two widely used architecture families: VGG and ResNet. Experiments on MNIST, CIFAR-10 and Imagenette show that, with the proposed method, the training time of the models is more than halved without significantly impacting accuracy. The FLOPs reduction in the forward propagation during training ranges from 17.83% for VGG-11 to 83.74% for ResNet-152. As for the accuracy, the impact depends on the depth of the model and the decrease is between 0.26% and 2.38% for VGGs and between 0.4 and 3.2% for ResNets. These results demonstrate the effectiveness of the proposed technique in speeding up learning of CNNs. The technique will be especially useful in applications where fine-tuning or online training of convolutional models is required, for instance because data arrive sequentially.

Image registration is the process of finding the geometric transformation that, applied to the floating image, gives the registered image with the highest similarity to the reference image. Registering a pair of images involves the definition of a similarity function in terms of the parameters of the geometric transformation that allows the registration. This paper proposes to register a pair of images by iteratively maximizing the empirical mutual information through coordinate gradient descent. Hence, the registered image is obtained by applying a sequence of local affine transformations. Rather than adopting a uniformly spaced grid to select image blocks to locally register, as done by state-of-the-art techniques, this paper proposes a method which is similar in spirit to boosting strategies used in classification. In this work, a probability distribution over the pixels of the registered image is maintained. At each pixel, this distribution represents the probability that a local affine transformation of a block centered on this pixel should be computed to improve the similarity between the registered and the reference images. The distribution is updated iteratively during the registration process to move probability mass towards pixels unaffected by the estimated local transformation. The paper presents preliminary results by a qualitative evaluation on several pairs of medical images acquired by different sources.

This paper proposes to model an action as the output of a sequence of atomic Linear Time Invariant (LTI) systems. The sequence of LTI systems generating the action is modeled as a Markov chain , where a Hidden Markov Model (HMM) is used to model the transition from one atomic LTI system to another. In turn, the LTI systems are represented in terms of their Hankel matrices. For classification purposes, the parameters of a set of HMMs (one for each action class) are learned via a discriminative approach. This work proposes a novel method to learn the atomic LTI systems from training data , and analyzes in detail the action representation in terms of a sequence of Hankel matrices. Extensive evaluation of the proposed approach on two publicly available datasets demonstrates that the proposed method attains state-of-the-art accuracy in action classification from the 3D locations of body joints (skeleton).

We introduce a new problem domain for activity recognition: the analysis of children's social and communicative behaviors based on video and audio data. We specifically target interactions between children aged 1-2 years and an adult. Such interactions arise naturally in the diagnosis and treatment of developmental disorders such as autism. We introduce a new publicly-available dataset containing over 160 sessions of a 3-5 minute child-adult interaction. In each session, the adult examiner followed a semi-structured play interaction protocol which was designed to elicit a broad range of social behaviors. We identify the key technical challenges in analyzing these behaviors, and describe methods for decoding the interactions. We present experimental results that demonstrate the potential of the dataset to drive interesting research questions, and show preliminary results for multi-modal activity recognition.

Tracking-by-detection is a widely used paradigm for multi-person tracking but is affected by variations in crowd density, obstacles in the scene, varying illumination, human pose variation, scale changes, etc. We propose an improved tracking-by-detection framework for multi-person tracking where the appearance model is formulated as a template ensemble updated online given detections provided by a pedestrian detector. We employ a hierarchy of trackers to select the most effective tracking strategy and an algorithm to adapt the conditions for trackers' initialization and termination. Our formulation is online and does not require calibration information. In experiments with four pedestrian tracking benchmark datasets, our formulation attains accuracy that is comparable to, or better than, the state-of-the-art pedestrian trackers that must exploit calibration information and operate offline.

We propose a framework for querying a distributed database of video surveillance data in order to retrieve a set of likely paths of a person moving in the area under surveillance. In our framework, each camera of the surveillance system locally processes the data and stores video sequences in a storage unit and the metadata for each detected person in the distributed database. A pedestrian's path is formulated as a dynamic Bayesian network (DBN) to model the dependencies between subsequent observations of the person as he makes his way through the camera network. We propose a tool by which the analyst can pose queries about where a certain person appeared while moving in the site during a specified temporal window. The DBN is used in an algorithm that finds potentially relevant metadata records from the distributed databases and then assembles these into probable paths that the person took in the camera network. Finally, the system presents the analyst with the retrieved set of likely paths in ranked order. The computational complexity for our method is quadratic in the number of camera nodes and linear in the number of moving persons. Experiments were carried out on simulated data to test the system with large distributed databases and in a real setting in which six databases store the data from six video cameras. The simulations confirm that our method provides good results with varying numbers of cameras and persons moving in the network. In a real setting, the method reconstructs paths across the camera network with approximatively 75% accuracy at rank 1.

In this paper, a new system is presented to support the user in the face annotation task. Every time a photo sequence becomes available, the system analyses it to detect and cluster faces in set corresponding to the same person. We propose to model the problem of people re-identification in photos as a data association problem. In this way, the system takes advantage from the assumption that each person can appear at most once in each photo. We propose a fully automated method for grouping facial images, the method does not require any initialization neither a priori knowledge of the number of persons that are in the photo sequence. We compare the results obtained with our method and with standard clustering methods on three personal collections and on a publicly available dataset.