Yi Zhu

Zhenlin Xu, Yi Zhu, Tiffany Deng et al.

This paper presents novel benchmarks for evaluating vision-language models (VLMs) in zero-shot recognition, focusing on granularity and specificity. Although VLMs excel in tasks like image captioning, they face challenges in open-world settings. Our benchmarks test VLMs' consistency in understanding concepts across semantic granularity levels and their response to varying text specificity. Findings show that VLMs favor moderately fine-grained concepts and struggle with specificity, often misjudging texts that differ from their training data. Extensive evaluations reveal limitations in current VLMs, particularly in distinguishing between correct and subtly incorrect descriptions. While fine-tuning offers some improvements, it doesn't fully address these issues, highlighting the need for VLMs with enhanced generalization capabilities for real-world applications. This study provides insights into VLM limitations and suggests directions for developing more robust models.

Yi Zhu, Surya Koppisetti, Trang Tran et al. · uw

Audio deepfake detection (ADD) is crucial to combat the misuse of speech synthesized from generative AI models. Existing ADD models suffer from generalization issues, with a large performance discrepancy between in-domain and out-of-domain data. Moreover, the black-box nature of existing models limits their use in real-world scenarios, where explanations are required for model decisions. To alleviate these issues, we introduce a new ADD model that explicitly uses the StyleLInguistics Mismatch (SLIM) in fake speech to separate them from real speech. SLIM first employs self-supervised pretraining on only real samples to learn the style-linguistics dependency in the real class. The learned features are then used in complement with standard pretrained acoustic features (e.g., Wav2vec) to learn a classifier on the real and fake classes. When the feature encoders are frozen, SLIM outperforms benchmark methods on out-of-domain datasets while achieving competitive results on in-domain data. The features learned by SLIM allow us to quantify the (mis)match between style and linguistic content in a sample, hence facilitating an explanation of the model decision.

Han Wang, Yi Zhu, Ye Wang et al.

Clickbait, which aims to induce users with some surprising and even thrilling headlines for increasing click-through rates, permeates almost all online content publishers, such as news portals and social media. Recently, Large Language Models (LLMs) have emerged as a powerful instrument and achieved tremendous success in a series of NLP downstream tasks. However, it is not yet known whether LLMs can be served as a high-quality clickbait detection system. In this paper, we analyze the performance of LLMs in the few-shot and zero-shot scenarios on several English and Chinese benchmark datasets. Experimental results show that LLMs cannot achieve the best results compared to the state-of-the-art deep and fine-tuning PLMs methods. Different from human intuition, the experiments demonstrated that LLMs cannot make satisfied clickbait detection just by the headlines.

Haofei Kuang, Yi Zhu, Zhi Zhang et al.

Contrastive learning has revolutionized self-supervised image representation learning field, and recently been adapted to video domain. One of the greatest advantages of contrastive learning is that it allows us to flexibly define powerful loss objectives as long as we can find a reasonable way to formulate positive and negative samples to contrast. However, existing approaches rely heavily on the short-range spatiotemporal salience to form clip-level contrastive signals, thus limit themselves from using global context. In this paper, we propose a new video-level contrastive learning method based on segments to formulate positive pairs. Our formulation is able to capture global context in a video, thus robust to temporal content change. We also incorporate a temporal order regularization term to enforce the inherent sequential structure of videos. Extensive experiments show that our video-level contrastive learning framework (VCLR) is able to outperform previous state-of-the-arts on five video datasets for downstream action classification, action localization and video retrieval. Code is available at https://github.com/amazon-research/video-contrastive-learning.

Zhiqiang Tang, Yunhe Gao, Yi Zhu et al.

Traditional normalization techniques (e.g., Batch Normalization and Instance Normalization) generally and simplistically assume that training and test data follow the same distribution. As distribution shifts are inevitable in real-world applications, well-trained models with previous normalization methods can perform badly in new environments. Can we develop new normalization methods to improve generalization robustness under distribution shifts? In this paper, we answer the question by proposing CrossNorm and SelfNorm. CrossNorm exchanges channel-wise mean and variance between feature maps to enlarge training distribution, while SelfNorm uses attention to recalibrate the statistics to bridge gaps between training and test distributions. CrossNorm and SelfNorm can complement each other, though exploring different directions in statistics usage. Extensive experiments on different fields (vision and language), tasks (classification and segmentation), settings (supervised and semi-supervised), and distribution shift types (synthetic and natural) show the effectiveness. Code is available at https://github.com/amazon-research/crossnorm-selfnorm

Pengfei Yu, Dongming Shen, Silin Meng et al.

We present RPGBench, the first benchmark designed to evaluate large language models (LLMs) as text-based role-playing game (RPG) engines. RPGBench comprises two core tasks: Game Creation (GC) and Game Simulation (GS). In GC, an LLM must craft a valid and playable RPG world using a structured event-state representation, ensuring logical coherence and proper termination conditions. In GS, the LLM simulates interactive gameplay across multiple rounds while consistently updating states and enforcing game rules. To comprehensively assess performance, RPGBench integrates objective and subjective evaluation methodologies. Objective measures verify adherence to event mechanics and check variable updates without requiring human intervention. Subjective measures, such as content interestingness, action quality, and role-playing capability, are evaluated via an LLM-as-a-judge framework, where a strong LLM grades each candidate's outputs. Empirical results demonstrate that state-of-the-art LLMs can produce engaging stories but often struggle to implement consistent, verifiable game mechanics, particularly in long or complex scenarios. By combining structured, rule-based assessments with LLM-based judgments, RPGBench provides a new standard for evaluating how well LLMs can balance creativity, coherence, and complexity in text-based RPGs, opening avenues for more immersive and controllable interactive storytelling.

Shengju Qian, Hao Shao, Yi Zhu et al.

The transformer architectures, based on self-attention mechanism and convolution-free design, recently found superior performance and booming applications in computer vision. However, the discontinuous patch-wise tokenization process implicitly introduces jagged artifacts into attention maps, arising the traditional problem of aliasing for vision transformers. Aliasing effect occurs when discrete patterns are used to produce high frequency or continuous information, resulting in the indistinguishable distortions. Recent researches have found that modern convolution networks still suffer from this phenomenon. In this work, we analyze the uncharted problem of aliasing in vision transformer and explore to incorporate anti-aliasing properties. Specifically, we propose a plug-and-play Aliasing-Reduction Module(ARM) to alleviate the aforementioned issue. We investigate the effectiveness and generalization of the proposed method across multiple tasks and various vision transformer families. This lightweight design consistently attains a clear boost over several famous structures. Furthermore, our module also improves data efficiency and robustness of vision transformers.

Fangrui Zhu, Yi Zhu, Li Zhang et al.

Semantic segmentation is a challenging problem due to difficulties in modeling context in complex scenes and class confusions along boundaries. Most literature either focuses on context modeling or boundary refinement, which is less generalizable in open-world scenarios. In this work, we advocate a unified framework(UN-EPT) to segment objects by considering both context information and boundary artifacts. We first adapt a sparse sampling strategy to incorporate the transformer-based attention mechanism for efficient context modeling. In addition, a separate spatial branch is introduced to capture image details for boundary refinement. The whole model can be trained in an end-to-end manner. We demonstrate promising performance on three popular benchmarks for semantic segmentation with low memory footprint. Code will be released soon.

Anish Pimpley, Shuo Li, Anubha Srivastava et al.

Optimizing resource allocation for analytical workloads is vital for reducing costs of cloud-data services. At the same time, it is incredibly hard for users to allocate resources per query in serverless processing systems, and they frequently misallocate by orders of magnitude. Unfortunately, prior work focused on predicting peak allocation while ignoring aggressive trade-offs between resource allocation and run-time. Additionally, these methods fail to predict allocation for queries that have not been observed in the past. In this paper, we tackle both these problems. We introduce a system for optimal resource allocation that can predict performance with aggressive trade-offs, for both new and past observed queries. We introduce the notion of a performance characteristic curve (PCC) as a parameterized representation that can compactly capture the relationship between resources and performance. To tackle training data sparsity, we introduce a novel data augmentation technique to efficiently synthesize the entire PCC using a single run of the query. Lastly, we demonstrate the advantages of a constrained loss function coupled with GNNs, over traditional ML methods, for capturing the domain specific behavior through an extensive experimental evaluation over SCOPE big data workloads at Microsoft.

Xueqing Deng, Yi Zhu, Yuxin Tian et al.

Neural network-based semantic segmentation has achieved remarkable results when large amounts of annotated data are available, that is, in the supervised case. However, such data is expensive to collect and so methods have been developed to adapt models trained on related, often synthetic data for which labels are readily available. Current adaptation approaches do not consider the dependence of the generalization/transferability of these models on network architecture. In this paper, we perform neural architecture search (NAS) to provide architecture-level perspective and analysis for domain adaptation. We identify the optimization gap that exists when searching architectures for unsupervised domain adaptation which makes this NAS problem uniquely difficult. We propose bridging this gap by using maximum mean discrepancy and regional weighted entropy to estimate the accuracy metric. Experimental results on several widely adopted benchmarks show that our proposed AutoAdapt framework indeed discovers architectures that improve the performance of a number of existing adaptation techniques.

Yanyi Zhang, Xinyu Li, Chunhui Liu et al.

We introduce Video Transformer (VidTr) with separable-attention for video classification. Comparing with commonly used 3D networks, VidTr is able to aggregate spatio-temporal information via stacked attentions and provide better performance with higher efficiency. We first introduce the vanilla video transformer and show that transformer module is able to perform spatio-temporal modeling from raw pixels, but with heavy memory usage. We then present VidTr which reduces the memory cost by 3.3$\times$ while keeping the same performance. To further optimize the model, we propose the standard deviation based topK pooling for attention ($pool_{topK\_std}$), which reduces the computation by dropping non-informative features along temporal dimension. VidTr achieves state-of-the-art performance on five commonly used datasets with lower computational requirement, showing both the efficiency and effectiveness of our design. Finally, error analysis and visualization show that VidTr is especially good at predicting actions that require long-term temporal reasoning.

Yi Zhu, Xinyu Li, Chunhui Liu et al.

Video action recognition is one of the representative tasks for video understanding. Over the last decade, we have witnessed great advancements in video action recognition thanks to the emergence of deep learning. But we also encountered new challenges, including modeling long-range temporal information in videos, high computation costs, and incomparable results due to datasets and evaluation protocol variances. In this paper, we provide a comprehensive survey of over 200 existing papers on deep learning for video action recognition. We first introduce the 17 video action recognition datasets that influenced the design of models. Then we present video action recognition models in chronological order: starting with early attempts at adapting deep learning, then to the two-stream networks, followed by the adoption of 3D convolutional kernels, and finally to the recent compute-efficient models. In addition, we benchmark popular methods on several representative datasets and release code for reproducibility. In the end, we discuss open problems and shed light on opportunities for video action recognition to facilitate new research ideas.

Xueqing Deng, Yi Zhu, Yuxin Tian et al.

Land-cover classification using remote sensing imagery is an important Earth observation task. Recently, land cover classification has benefited from the development of fully connected neural networks for semantic segmentation. The benchmark datasets available for training deep segmentation models in remote sensing imagery tend to be small, however, often consisting of only a handful of images from a single location with a single scale. This limits the models' ability to generalize to other datasets. Domain adaptation has been proposed to improve the models' generalization but we find these approaches are not effective for dealing with the scale variation commonly found between remote sensing image collections. We therefore propose a scale aware adversarial learning framework to perform joint cross-location and cross-scale land-cover classification. The framework has a dual discriminator architecture with a standard feature discriminator as well as a novel scale discriminator. We also introduce a scale attention module which produces scale-enhanced features. Experimental results show that the proposed framework outperforms state-of-the-art domain adaptation methods by a large margin.

Yi Zhu, Zhongyue Zhang, Chongruo Wu et al.

Deep learning usually achieves the best results with complete supervision. In the case of semantic segmentation, this means that large amounts of pixelwise annotations are required to learn accurate models. In this paper, we show that we can obtain state-of-the-art results using a semi-supervised approach, specifically a self-training paradigm. We first train a teacher model on labeled data, and then generate pseudo labels on a large set of unlabeled data. Our robust training framework can digest human-annotated and pseudo labels jointly and achieve top performances on Cityscapes, CamVid and KITTI datasets while requiring significantly less supervision. We also demonstrate the effectiveness of self-training on a challenging cross-domain generalization task, outperforming conventional finetuning method by a large margin. Lastly, to alleviate the computational burden caused by the large amount of pseudo labels, we propose a fast training schedule to accelerate the training of segmentation models by up to 2x without performance degradation.

Hu Zhang, Linchao Zhu, Yi Zhu et al.

Deep neural networks are known to be susceptible to adversarial noise, which are tiny and imperceptible perturbations. Most of previous work on adversarial attack mainly focus on image models, while the vulnerability of video models is less explored. In this paper, we aim to attack video models by utilizing intrinsic movement pattern and regional relative motion among video frames. We propose an effective motion-excited sampler to obtain motion-aware noise prior, which we term as sparked prior. Our sparked prior underlines frame correlations and utilizes video dynamics via relative motion. By using the sparked prior in gradient estimation, we can successfully attack a variety of video classification models with fewer number of queries. Extensive experimental results on four benchmark datasets validate the efficacy of our proposed method.

Xueqing Deng, Yi Zhu, Yuxin Tian et al.

That most deep learning models are purely data driven is both a strength and a weakness. Given sufficient training data, the optimal model for a particular problem can be learned. However, this is usually not the case and so instead the model is either learned from scratch from a limited amount of training data or pre-trained on a different problem and then fine-tuned. Both of these situations are potentially suboptimal and limit the generalizability of the model. Inspired by this, we investigate methods to inform or guide deep learning models for geospatial image analysis to increase their performance when a limited amount of training data is available or when they are applied to scenarios other than which they were trained on. In particular, we exploit the fact that there are certain fundamental rules as to how things are distributed on the surface of the Earth and these rules do not vary substantially between locations. Based on this, we develop a novel feature pooling method for convolutional neural networks using Getis-Ord Gi* analysis from geostatistics. Experimental results show our proposed pooling function has significantly better generalization performance compared to a standard data-driven approach when applied to overhead image segmentation.

Yi Zhu, Shawn Newsam

Motivated by our observation that motion information is the key to good anomaly detection performance in video, we propose a temporal augmented network to learn a motion-aware feature. This feature alone can achieve competitive performance with previous state-of-the-art methods, and when combined with them, can achieve significant performance improvements. Furthermore, we incorporate temporal context into the Multiple Instance Learning (MIL) ranking model by using an attention block. The learned attention weights can help to differentiate between anomalous and normal video segments better. With the proposed motion-aware feature and the temporal MIL ranking model, we outperform previous approaches by a large margin on both anomaly detection and anomalous action recognition tasks in the UCF Crime dataset.

Yi Zhu

In this dissertation, I present my work towards exploring temporal information for better video understanding. Specifically, I have worked on two problems: action recognition and semantic segmentation. For action recognition, I have proposed a framework, termed hidden two-stream networks, to learn an optimal motion representation that does not require the computation of optical flow. My framework alleviates several challenges faced in video classification, such as learning motion representations, real-time inference, multi-framerate handling, generalizability to unseen actions, etc. For semantic segmentation, I have introduced a general framework that uses video prediction models to synthesize new training samples. By scaling up the training dataset, my trained models are more accurate and robust than previous models even without modifications to the network architectures or objective functions. I believe videos have much more potential to be mined, and temporal information is one of the most important cues for machines to perceive the visual world better.

Xueqing Deng, Yi Zhu, Shawn Newsam

This paper develops a deep-learning framework to synthesize a ground-level view of a location given an overhead image. We propose a novel conditional generative adversarial network (cGAN) in which the trained generator generates realistic looking and representative ground-level images using overhead imagery as auxiliary information. The generator is an encoder-decoder network which allows us to compare low- and high-level features as well as their concatenation for encoding the overhead imagery. We also demonstrate how our framework can be used to perform land cover classification by modifying the trained cGAN to extract features from overhead imagery. This is interesting because, although we are using this modified cGAN as a feature extractor for overhead imagery, it incorporates knowledge of how locations look from the ground.

Yi Zhu, Karan Sapra, Fitsum A. Reda et al.

Semantic segmentation requires large amounts of pixel-wise annotations to learn accurate models. In this paper, we present a video prediction-based methodology to scale up training sets by synthesizing new training samples in order to improve the accuracy of semantic segmentation networks. We exploit video prediction models' ability to predict future frames in order to also predict future labels. A joint propagation strategy is also proposed to alleviate mis-alignments in synthesized samples. We demonstrate that training segmentation models on datasets augmented by the synthesized samples leads to significant improvements in accuracy. Furthermore, we introduce a novel boundary label relaxation technique that makes training robust to annotation noise and propagation artifacts along object boundaries. Our proposed methods achieve state-of-the-art mIoUs of 83.5% on Cityscapes and 82.9% on CamVid. Our single model, without model ensembles, achieves 72.8% mIoU on the KITTI semantic segmentation test set, which surpasses the winning entry of the ROB challenge 2018. Our code and videos can be found at https://nv-adlr.github.io/publication/2018-Segmentation.

Yi Zhu, Shawn Newsam

Current state-of-the-art approaches to video understanding adopt temporal jittering to simulate analyzing the video at varying frame rates. However, this does not work well for multirate videos, in which actions or subactions occur at different speeds. The frame sampling rate should vary in accordance with the different motion speeds. In this work, we propose a simple yet effective strategy, termed random temporal skipping, to address this situation. This strategy effectively handles multirate videos by randomizing the sampling rate during training. It is an exhaustive approach, which can potentially cover all motion speed variations. Furthermore, due to the large temporal skipping, our network can see video clips that originally cover over 100 frames. Such a time range is enough to analyze most actions/events. We also introduce an occlusion-aware optical flow learning method that generates improved motion maps for human action recognition. Our framework is end-to-end trainable, runs in real-time, and achieves state-of-the-art performance on six widely adopted video benchmarks.

Yi Zhu, Jia Xue, Shawn Newsam

Deep neural networks have led to a series of breakthroughs in computer vision given sufficient annotated training datasets. For novel tasks with limited labeled data, the prevalent approach is to transfer the knowledge learned in the pre-trained models to the new tasks by fine-tuning. Classic model fine-tuning utilizes the fact that well trained neural networks appear to learn cross domain features. These features are treated equally during transfer learning. In this paper, we explore the impact of feature selection in model fine-tuning by introducing a transfer module, which assigns weights to features extracted from pre-trained models. The proposed transfer module proves the importance of feature selection for transferring models from source to target domains. It is shown to significantly improve upon fine-tuning results with only marginal extra computational cost. We also incorporate an auxiliary classifier as an extra regularizer to avoid over-fitting. Finally, we build a Gated Transfer Network (GTN) based on our transfer module and achieve state-of-the-art results on six different tasks.

Xueqing Deng, Yi Zhu, Shawn Newsam

This paper investigates conditional generative adversarial networks (cGANs) to overcome a fundamental limitation of using geotagged media for geographic discovery, namely its sparse and uneven spatial distribution. We train a cGAN to generate ground-level views of a location given overhead imagery. We show the "fake" ground-level images are natural looking and are structurally similar to the real images. More significantly, we show the generated images are representative of the locations and that the representations learned by the cGANs are informative. In particular, we show that dense feature maps generated using our framework are more effective for land-cover classification than approaches which spatially interpolate features extracted from sparse ground-level images. To our knowledge, ours is the first work to use cGANs to generate ground-level views given overhead imagery and to explore the benefits of the learned representations.

Yi Zhu, Shawn Newsam

Despite the significant progress that has been made on estimating optical flow recently, most estimation methods, including classical and deep learning approaches, still have difficulty with multi-scale estimation, real-time computation, and/or occlusion reasoning. In this paper, we introduce dilated convolution and occlusion reasoning into unsupervised optical flow estimation to address these issues. The dilated convolution allows our network to avoid upsampling via deconvolution and the resulting gridding artifacts. Dilated convolution also results in a smaller memory footprint which speeds up interference. The occlusion reasoning prevents our network from learning incorrect deformations due to occluded image regions during training. Our proposed method outperforms state-of-the-art unsupervised approaches on the KITTI benchmark. We also demonstrate its generalization capability by applying it to action recognition in video.

Yanzhao Zhou, Yi Zhu, Qixiang Ye et al.

Weakly supervised instance segmentation with image-level labels, instead of expensive pixel-level masks, remains unexplored. In this paper, we tackle this challenging problem by exploiting class peak responses to enable a classification network for instance mask extraction. With image labels supervision only, CNN classifiers in a fully convolutional manner can produce class response maps, which specify classification confidence at each image location. We observed that local maximums, i.e., peaks, in a class response map typically correspond to strong visual cues residing inside each instance. Motivated by this, we first design a process to stimulate peaks to emerge from a class response map. The emerged peaks are then back-propagated and effectively mapped to highly informative regions of each object instance, such as instance boundaries. We refer to the above maps generated from class peak responses as Peak Response Maps (PRMs). PRMs provide a fine-detailed instance-level representation, which allows instance masks to be extracted even with some off-the-shelf methods. To the best of our knowledge, we for the first time report results for the challenging image-level supervised instance segmentation task. Extensive experiments show that our method also boosts weakly supervised pointwise localization as well as semantic segmentation performance, and reports state-of-the-art results on popular benchmarks, including PASCAL VOC 2012 and MS COCO.

Yi Zhu, Yang Long, Yu Guan et al.

Unseen Action Recognition (UAR) aims to recognise novel action categories without training examples. While previous methods focus on inner-dataset seen/unseen splits, this paper proposes a pipeline using a large-scale training source to achieve a Universal Representation (UR) that can generalise to a more realistic Cross-Dataset UAR (CD-UAR) scenario. We first address UAR as a Generalised Multiple-Instance Learning (GMIL) problem and discover 'building-blocks' from the large-scale ActivityNet dataset using distribution kernels. Essential visual and semantic components are preserved in a shared space to achieve the UR that can efficiently generalise to new datasets. Predicted UR exemplars can be improved by a simple semantic adaptation, and then an unseen action can be directly recognised using UR during the test. Without further training, extensive experiments manifest significant improvements over the UCF101 and HMDB51 benchmarks.

Xueqing Deng, Yi Zhu, Shawn Newsam

In recent years, geotagged social media has become popular as a novel source for geographic knowledge discovery. Ground-level images and videos provide a different perspective than overhead imagery and can be applied to a range of applications such as land use mapping, activity detection, pollution mapping, etc. The sparse and uneven distribution of this data presents a problem, however, for generating dense maps. We therefore investigate the problem of spatially interpolating the high-dimensional features extracted from sparse social media to enable dense labeling using standard classifiers. Further, we show how prior knowledge about region boundaries can be used to improve the interpolation through spatial morphing kernel regression. We show that an interpolate-then-classify framework can produce dense maps from sparse observations but that care must be taken in choosing the interpolation method. We also show that the spatial morphing kernel improves the results.

Yi Zhu, Xueqing Deng, Shawn Newsam

We perform fine-grained land use mapping at the city scale using ground-level images. Mapping land use is considerably more difficult than mapping land cover and is generally not possible using overhead imagery as it requires close-up views and seeing inside buildings. We postulate that the growing collections of georeferenced, ground-level images suggest an alternate approach to this geographic knowledge discovery problem. We develop a general framework that uses Flickr images to map 45 different land-use classes for the City of San Francisco. Individual images are classified using a novel convolutional neural network containing two streams, one for recognizing objects and another for recognizing scenes. This network is trained in an end-to-end manner directly on the labeled training images. We propose several strategies to overcome the noisiness of our user-generated data including search-based training set augmentation and online adaptive training. We derive a ground truth map of San Francisco in order to evaluate our method. We demonstrate the effectiveness of our approach through geo-visualization and quantitative analysis. Our framework achieves over 29% recall at the individual land parcel level which represents a strong baseline for the challenging 45-way land use classification problem especially given the noisiness of the image data.

Yi Zhu, Yanzhao Zhou, Qixiang Ye et al.

Weakly supervised object localization remains challenging, where only image labels instead of bounding boxes are available during training. Object proposal is an effective component in localization, but often computationally expensive and incapable of joint optimization with some of the remaining modules. In this paper, to the best of our knowledge, we for the first time integrate weakly supervised object proposal into convolutional neural networks (CNNs) in an end-to-end learning manner. We design a network component, Soft Proposal (SP), to be plugged into any standard convolutional architecture to introduce the nearly cost-free object proposal, orders of magnitude faster than state-of-the-art methods. In the SP-augmented CNNs, referred to as Soft Proposal Networks (SPNs), iteratively evolved object proposals are generated based on the deep feature maps then projected back, and further jointly optimized with network parameters, with image-level supervision only. Through the unified learning process, SPNs learn better object-centric filters, discover more discriminative visual evidence, and suppress background interference, significantly boosting both weakly supervised object localization and classification performance. We report the best results on popular benchmarks, including PASCAL VOC, MS COCO, and ImageNet.

Yi Zhu, Shawn Newsam

Classical approaches for estimating optical flow have achieved rapid progress in the last decade. However, most of them are too slow to be applied in real-time video analysis. Due to the great success of deep learning, recent work has focused on using CNNs to solve such dense prediction problems. In this paper, we investigate a new deep architecture, Densely Connected Convolutional Networks (DenseNet), to learn optical flow. This specific architecture is ideal for the problem at hand as it provides shortcut connections throughout the network, which leads to implicit deep supervision. We extend current DenseNet to a fully convolutional network to learn motion estimation in an unsupervised manner. Evaluation results on three standard benchmarks demonstrate that DenseNet is a better fit than other widely adopted CNN architectures for optical flow estimation.

Yi Zhu, Sen Liu, Shawn Newsam

This paper is the first work to perform spatio-temporal mapping of human activity using the visual content of geo-tagged videos. We utilize a recent deep-learning based video analysis framework, termed hidden two-stream networks, to recognize a range of activities in YouTube videos. This framework is efficient and can run in real time or faster which is important for recognizing events as they occur in streaming video or for reducing latency in analyzing already captured video. This is, in turn, important for using video in smart-city applications. We perform a series of experiments to show our approach is able to accurately map activities both spatially and temporally. We also demonstrate the advantages of using the visual content over the tags/titles.

Yi Zhu, Shawn Newsam, Zaikun Xu

This notebook paper describes our system for the untrimmed classification task in the ActivityNet challenge 2016. We investigate multiple state-of-the-art approaches for action recognition in long, untrimmed videos. We exploit hand-crafted motion boundary histogram features as well feature activations from deep networks such as VGG16, GoogLeNet, and C3D. These features are separately fed to linear, one-versus-rest support vector machine classifiers to produce confidence scores for each action class. These predictions are then fused along with the softmax scores of the recent ultra-deep ResNet-101 using weighted averaging.

Yi Zhu, Zhenzhong Lan, Shawn Newsam et al.

Analyzing videos of human actions involves understanding the temporal relationships among video frames. State-of-the-art action recognition approaches rely on traditional optical flow estimation methods to pre-compute motion information for CNNs. Such a two-stage approach is computationally expensive, storage demanding, and not end-to-end trainable. In this paper, we present a novel CNN architecture that implicitly captures motion information between adjacent frames. We name our approach hidden two-stream CNNs because it only takes raw video frames as input and directly predicts action classes without explicitly computing optical flow. Our end-to-end approach is 10x faster than its two-stage baseline. Experimental results on four challenging action recognition datasets: UCF101, HMDB51, THUMOS14 and ActivityNet v1.2 show that our approach significantly outperforms the previous best real-time approaches.

Yi Zhu, Zhenzhong Lan, Shawn Newsam et al.

We study the unsupervised learning of CNNs for optical flow estimation using proxy ground truth data. Supervised CNNs, due to their immense learning capacity, have shown superior performance on a range of computer vision problems including optical flow prediction. They however require the ground truth flow which is usually not accessible except on limited synthetic data. Without the guidance of ground truth optical flow, unsupervised CNNs often perform worse as they are naturally ill-conditioned. We therefore propose a novel framework in which proxy ground truth data generated from classical approaches is used to guide the CNN learning. The models are further refined in an unsupervised fashion using an image reconstruction loss. Our guided learning approach is competitive with or superior to state-of-the-art approaches on three standard benchmark datasets yet is completely unsupervised and can run in real time.

Zhenzhong Lan, Yi Zhu, Alexander G. Hauptmann

We investigate the problem of representing an entire video using CNN features for human action recognition. Currently, limited by GPU memory, we have not been able to feed a whole video into CNN/RNNs for end-to-end learning. A common practice is to use sampled frames as inputs and video labels as supervision. One major problem of this popular approach is that the local samples may not contain the information indicated by global labels. To deal with this problem, we propose to treat the deep networks trained on local inputs as local feature extractors. After extracting local features, we aggregate them into global features and train another mapping function on the same training data to map the global features into global labels. We study a set of problems regarding this new type of local features such as how to aggregate them into global features. Experimental results on HMDB51 and UCF101 datasets show that, for these new local features, a simple maximum pooling on the sparsely sampled features lead to significant performance improvement.

Yi Zhu, Shawn Newsam

This paper studies the joint learning of action recognition and temporal localization in long, untrimmed videos. We employ a multi-task learning framework that performs the three highly related steps of action proposal, action recognition, and action localization refinement in parallel instead of the standard sequential pipeline that performs the steps in order. We develop a novel temporal actionness regression module that estimates what proportion of a clip contains action. We use it for temporal localization but it could have other applications like video retrieval, surveillance, summarization, etc. We also introduce random shear augmentation during training to simulate viewpoint change. We evaluate our framework on three popular video benchmarks. Results demonstrate that our joint model is efficient in terms of storage and computation in that we do not need to compute and cache dense trajectory features, and that it is several times faster than its sequential ConvNets counterpart. Yet, despite being more efficient, it outperforms state-of-the-art methods with respect to accuracy.

Yi Zhu, Shawn Newsam

We perform spatio-temporal analysis of public sentiment using geotagged photo collections. We develop a deep learning-based classifier that predicts the emotion conveyed by an image. This allows us to associate sentiment with place. We perform spatial hotspot detection and show that different emotions have distinct spatial distributions that match expectations. We also perform temporal analysis using the capture time of the photos. Our spatio-temporal hotspot detection correctly identifies emerging concentrations of specific emotions and year-by-year analyses of select locations show there are strong temporal correlations between the predicted emotions and known events.

Yi Zhu, Shawn Newsam

Land use mapping is a fundamental yet challenging task in geographic science. In contrast to land cover mapping, it is generally not possible using overhead imagery. The recent, explosive growth of online geo-referenced photo collections suggests an alternate approach to geographic knowledge discovery. In this work, we present a general framework that uses ground-level images from Flickr for land use mapping. Our approach benefits from several novel aspects. First, we address the nosiness of the online photo collections, such as imprecise geolocation and uneven spatial distribution, by performing location and indoor/outdoor filtering, and semi- supervised dataset augmentation. Our indoor/outdoor classifier achieves state-of-the-art performance on several bench- mark datasets and approaches human-level accuracy. Second, we utilize high-level semantic image features extracted using deep learning, specifically convolutional neural net- works, which allow us to achieve upwards of 76% accuracy on a challenging eight class land use mapping problem.

Yi Zhu, Shawn Newsam

This paper performs the first investigation into depth for large-scale human action recognition in video where the depth cues are estimated from the videos themselves. We develop a new framework called depth2action and experiment thoroughly into how best to incorporate the depth information. We introduce spatio-temporal depth normalization (STDN) to enforce temporal consistency in our estimated depth sequences. We also propose modified depth motion maps (MDMM) to capture the subtle temporal changes in depth. These two components significantly improve the action recognition performance. We evaluate our depth2action framework on three large-scale action recognition video benchmarks. Our model achieves state-of-the-art performance when combined with appearance and motion information thus demonstrating that depth2action is indeed complementary to existing approaches.