Code translation aims to convert source code from one programming language (PL) to another. Given the promising abilities of large language models (LLMs) in code synthesis, researchers are exploring their potential to automate code translation. The prerequisite for advancing the state of LLM-based code translation is to understand their promises and limitations over existing techniques. To that end, we present a large-scale empirical study to investigate the ability of general LLMs and code LLMs for code translation across pairs of different languages, including C, C++, Go, Java, and Python. Our study, which involves the translation of 1,700 code samples from three benchmarks and two real-world projects, reveals that LLMs are yet to be reliably used to automate code translation---with correct translations ranging from 2.1% to 47.3% for the studied LLMs. Further manual investigation of unsuccessful translations identifies 15 categories of translation bugs. We also compare LLM-based code translation with traditional non-LLM-based approaches. Our analysis shows that these two classes of techniques have their own strengths and weaknesses. Finally, insights from our study suggest that providing more context to LLMs during translation can help them produce better results. To that end, we propose a prompt-crafting approach based on the symptoms of erroneous translations; this improves the performance of LLM-based code translation by 5.5% on average. Our study is the first of its kind, in terms of scale and breadth, that provides insights into the current limitations of LLMs in code translation and opportunities for improving them. Our dataset---consisting of 1,700 code samples in five PLs with 10K+ tests, 43K+ translated code, 1,748 manually labeled bugs, and 1,365 bug-fix pairs---can help drive research in this area.

Recent progress on vision and language research has shown that visual supervision improves the performance of large language models (LLMs) in various natural language processing (NLP) tasks. In particular, the Vokenization approach [65] initiated a new way of incorporating visual information into LLM training, demonstrating the potential of visual supervision for NLP tasks in a monolingual (i.e., English) setting. Given the effectiveness of visual information in human communication among people who speak different languages, we tackle an ambitious question in this paper; can we expect that visual supervision contributes to cross-lingual transfer learning from a high-resource language to low-resource languages in NLP tasks? To study this hypothesis, we build a cross-lingual Vokenization model and train a cross-lingual LLM on three languages, English, Urdu, and Swahili, in which the last two are considered low-resource languages. The experimental results demonstrate that our visually-supervised cross-lingual transfer learning method significantly improves the LLM performance in multiple cross-lingual NLP tasks such as XNLI, NER, and TyDiQA tasks for low-resource languages. We also qualitatively and quantitatively demonstrate that the benefit of our approach increases as the linguistic distance between low-and high-resource languages grows larger.

Neural Architecture Search (NAS) is an open and challenging problem in machine learning. While NAS offers great promise, the prohibitive computational demand of most of the existing NAS methods makes it difficult to directly search the architectures on large-scale tasks. The typical way of conducting large scale NAS is to search for an architectural building block on a small dataset (either using a proxy set from the large dataset or a completely different small scale dataset) and then transfer the block to a larger dataset. Despite a number of recent results that show the promise of transfer from proxy datasets, a comprehensive evaluation of different NAS methods studying the impact of different source datasets has not yet been addressed. In this work, we propose to analyze the architecture transferability of different NAS methods by performing a series of experiments on large scale benchmarks such as ImageNet1K and ImageNet22K. We find that: (i) The size and domain of the proxy set does not seem to influence architecture performance on the target dataset. On average, transfer performance of architectures searched using completely different small datasets (e.g., CIFAR10) perform similarly to the architectures searched directly on proxy target datasets. However, design of proxy sets has considerable impact on rankings of different NAS methods. (ii) While different NAS methods show similar performance on a source dataset (e.g., CIFAR10), they significantly differ on the transfer performance to a large dataset (e.g., ImageNet1K). (iii) Even on large datasets, random sampling baseline is very competitive, but the choice of the appropriate combination of proxy set and search strategy can provide significant improvement over it. We believe that our extensive empirical analysis will prove useful for future design of NAS algorithms.

The production of sports highlight packages summarizing a game's most exciting moments is an essential task for broadcast media. Yet, it requires labor-intensive video editing. We propose a novel approach for auto-curating sports highlights, and demonstrate it to create a first of a kind, real-world system for the editorial aid of golf and tennis highlight reels. Our method fuses information from the players’ reactions (action recognition such as high-fives and fist pumps), players’ expressions (aggressive, tense, smiling, and neutral), spectators (crowd cheering), commentator (tone of the voice and word analysis), and game analytics to determine the most interesting moments of a game. We accurately identify the start and end frames of key shot highlights with additional metadata, such as the player's name and the whole number, or analysts input allowing personalized content summarization and retrieval. In addition, we introduce new techniques for learning our classifiers with reduced manual training data annotation by exploiting the correlation of different modalities. Our work has been demonstrated at a major golf tournament (2017 Masters) and two major international tennis tournaments (2017 Wimbledon and U.S. Open), successfully extracting highlights through the course of the sporting events. For the 2017 Masters, 54% of the clips selected by our system overlapped with the official highlights reels. Furthermore, user studies showed that 90% of the non-overlapping ones were of the same quality of the official clips for the 2017 Masters, while the automatic selection of clips for highlights of 2017 Wimbledon and 2017 US Open agreed with human preferences 80% and 84.2% of the time, respectively.

We introduce a novel multi-modal system for auto-curating golf highlights that fuses information from players' reactions (celebration actions), spectators (crowd cheering), and commentator (tone of the voice and word analysis) to determine the most interesting moments of a game. The start of a highlight is determined with additional metadata (player's name and the hole number), allowing personalized content summarization and retrieval. Our system was demonstrated at Masters 2017, a major golf tournament, generating real-time highlights from four live video streams over four days.

We propose a visual food recognition framework that integrates the inherent semantic relationships among fine-grained classes. Our method learns semantics-aware features by formulating a multi-task loss function on top of a convolutional neural network (CNN) architecture. It then refines the CNN predictions using a random walk based smoothing procedure, which further exploits the rich semantic information. We evaluate our algorithm on a large "food-in-the-wild" benchmark, as well as a challenging dataset of restaurant food dishes with very few training images. The proposed method achieves higher classification accuracy than a baseline which directly fine-tunes a deep learning network on the target dataset. Furthermore, we analyze the consistency of the learned model with the inherent semantic relationships among food categories. Results show that the proposed approach provides more semantically meaningful results than the baseline method, even in cases of mispredictions.

We propose a method to extract user attributes from the pictures posted in social media feeds, specifically gender information. While traditional approaches rely on text analysis or exploit visual information only from the user profile picture or colors, we propose to look at the distribution of semantics in the pictures coming from the whole feed of a person to estimate gender. In order to compute such semantic distribution, we trained models from existing visual taxonomies to recognize objects, scenes and activities, and applied them to the images in each user's feed. Experiments conducted on a set of ten thousand twitter users and their collection of half a million images revealed that the gender signal can indeed be extracted from the users image feed (75.6% accuracy). Furthermore, the combination of visual cues resulted almost as strong as textual analysis in predicting gender, while providing complementary information that can be employed to further boost gender prediction accuracy to 88% when combined with textual data. As a byproduct of our investigation, we were also able to extrapolate the semantic categories of posted pictures mostly correlated to males and females.

Action recognition is an important problem in computer vision and has received substantial attention in recent years. However, it remains very challenging due to the complex interaction of static and dynamic information, as well as the high computational cost of processing video data. This paper aims to apply the success of static image semantic recognition to the video domain, by leveraging both static and motion based descriptors in different stages of the semantic ladder. We examine the effects of three types of features: low-level dynamic descriptors, intermediate-level static deep architecture outputs, and static high-level semantics. In order to combine such heterogeneous sources of information, we employ a scalable method to fuse these features. Through extensive experimental evaluations, we demonstrate that the proposed framework significantly improves action classification performance. We have obtained an accuracy of 89.59% and 62.88% on the well-known UCF-101 and HMDB-51 benchmarks, respectively, which compare favorably with the state-of-the-art.

Attribute-based representation has been widely used in visual recognition and retrieval due to its interpretability and cross-category generalization properties. However, classic attribute learning requires manually labeling attributes on the images, which is very expensive, and not scalable. In this paper, we propose to model attributes from category-attribute proportions. The proposed framework can model attributes without attribute labels on the images. Specifically, given a multi-class image datasets with N categories, we model an attribute, based on an N-dimensional category-attribute proportion vector, where each element of the vector characterizes the proportion of images in the corresponding category having the attribute. The attribute learning can be formulated as a learning from label proportion (LLP) problem. Our method is based on a newly proposed machine learning algorithm called $\propto$SVM. Finding the category-attribute proportions is much easier than manually labeling images, but it is still not a trivial task. We further propose to estimate the proportions from multiple modalities such as human commonsense knowledge, NLP tools, and other domain knowledge. The value of the proposed approach is demonstrated by various applications including modeling animal attributes, visual sentiment attributes, and scene attributes.

Video event classification and retrieval has recently emerged as a challenging research topic. In addition to the variation in appearance of visual content and the large scale of the collections to be analyzed, this domain presents new and unique challenges in the modeling of the explicit temporal structure and implicit temporal trends of content within the video events. In this study, we present a technique for video event classification that captures temporal information over semantics using a scalable and efficient modeling scheme. An architecture for partitioning videos into a linear temporal pyramid, using segments of equal length and segments determined by the patterns of the underlying data, is applied over a rich underlying semantic description at the frame level using a taxonomy of nearly 1000 concepts containing 500,000 training images. Forward model selection with data bagging is used to prune the space of temporal features and data for efficiency. The system is implemented in the Hadoop Map-Reduce environment for arbitrary scalability. Our method is applied to the TRECVID Multimedia Event Detection 2012 task. Results demonstrate a significant boost in performance of over 50%, in terms of mean average precision, compared to common max or average pooling, and 17.7% compared to more complex pooling strategies that ignore temporal content.

We propose semantic model vectors, an intermediate level semantic representation, as a basis for modeling and detecting complex events in unconstrained real-world videos, such as those from YouTube. The semantic model vectors are extracted using a set of discriminative semantic classifiers, each being an ensemble of SVM models trained from thousands of labeled web images, for a total of 280 generic concepts. Our study reveals that the proposed semantic model vectors representation outperforms-and is complementary to-other low-level visual descriptors for video event modeling. We hence present an end-to-end video event detection system, which combines semantic model vectors with other static or dynamic visual descriptors, extracted at the frame, segment, or full clip level. We perform a comprehensive empirical study on the 2010 TRECVID Multimedia Event Detection task (http://www.nist.gov/itl/iad/mig/med10.cfm), which validates the semantic model vectors representation not only as the best individual descriptor, outperforming state-of-the-art global and local static features as well as spatio-temporal HOG and HOF descriptors, but also as the most compact. We also study early and late feature fusion across the various approaches, leading to a 15% performance boost and an overall system performance of 0.46 mean average precision. In order to promote further research in this direction, we made our semantic model vectors for the TRECVID MED 2010 set publicly available for the community to use (http://www1.cs.columbia.edu/~mmerler/SMV.html).

The exponential diffusion of unstructured multimedia content on sites such as YouTube has lately fostered a rapid growth of interest in the multimedia community toward a new line of systems to recognize visual content "in the wild", that is, in less structured, unconstrained and more realistic domains. Due to the lack of structure and to the low quality of the data, algorithms and paradigms designed for professional content often cannot directly be applied to the aforementioned domains, thus presenting a new challenge.

We propose a system to select the most representative faces in unstructured presentation videos with respect to two criteria: to optimize matching accuracy between pairs of face tracks, and to select humanly preferred face icons for indexing purposes. We first extract face tracks using state-of-the-art face detection and tracking. A small subset of images are then selected per track in order to maximize matching accuracy between tracks. Finally, representative images are extracted for each speaker in order to build a face index of the video. We tested our approach on 3 unstructured presentation videos of approximately 45 minutes each, for a total of a quarter million frames. Compared to the standard min-min approach, our method achieves higher track matching accuracy (94.22%), while using 6% of the running time. Using an optimal combination of 3 user preference measures, we were able to build face indexes containing 54 speakers (out of the 58 present in the videos) indexing into 795 detected tracks.

Ranking large scale image and video collections usually expects higher accuracy on top ranked data, while tolerates lower accuracy on bottom ranked ones. In view of this, we propose a rank learning algorithm, called Imbalanced RankBoost, which merges RankBoost and iterative thresholding into a unified loss optimization framework. The proposed approach provides a more efficient ranking process by iteratively identifying a cutoff threshold in each boosting iteration, and automatically truncating ranking feature computation for the data ranked below. Experiments on the TRECVID 2007 high-level feature benchmark show that the proposed approach outperforms RankBoost in terms of both ranking effectiveness and efficiency. It achieves an up to 21% improvement in terms of mean average precision, or equivalently, a 6-fold speedup in the ranking process.

We propose a fully automatic method for summarizing and indexing unstructured presentation videos based on text extracted from the projected slides. We use changes of text in the slides as a means to segment the video into semantic shots. Unlike precedent approaches, our method does not depend on availability of the electronic source of the slides, but rather extracts and recognizes the text directly from the video. Once text regions are detected within keyframes, a novel binarization algorithm, Local Adaptive Otsu (LOA), is employed to deal with the low quality of video scene text, before feeding the regions to the open source Tesseract1OCR engine for recognition. We tested our system on a corpus of 8 presentation videos for a total of 1 hour and 45 minutes, achieving 0.5343 Precision and 0.7446 Recall Character recognition rates, and 0.4947 Precision and 0.6651 Recall Word recognition rates. Besides being used for multimedia documents, topic indexing, and cross referencing, our system can be integrated into summarization and presentation tools such as the VAST MultiMedia browser.

The problem of using pictures of objects captured under ideal imaging conditions (here referred to as in vitro) to recognize objects in natural environments (in situ) is an emerging area of interest in computer vision and pattern recognition. Examples of tasks in this vein include assistive vision systems for the blind and object recognition for mobile robots; the proliferation of image databases on the web is bound to lead to more examples in the near future. Despite its importance, there is still a need for a freely available database to facilitate study of this kind of training/testing dichotomy. In this work one of our contributions is a new multimedia database of 120 grocery products, GroZi-120. For every product, two different recordings are available: in vitro images extracted from the web, and in situ images extracted from camcorder video collected inside a grocery store. As an additional contribution, we present the results of applying three commonly used object recognition/detection algorithms (color histogram matching, SIFT matching, and boosted Haar-like features) to the dataset. Finally, we analyze the successes and failures of these algorithms against product type and imaging conditions, both in terms of recognition rate and localization accuracy, in order to suggest ways forward for further research in this domain.