Thomas Huang

Hyunjae Kim, Jiwoong Sohn, Aidan Gilson et al.

Large language models (LLMs) are transforming the landscape of medicine, yet two fundamental challenges persist: keeping up with rapidly evolving medical knowledge and providing verifiable, evidence-grounded reasoning. Retrieval-augmented generation (RAG) has been widely adopted to address these limitations by supplementing model outputs with retrieved evidence. However, whether RAG reliably achieves these goals remains unclear. Here, we present the most comprehensive expert evaluation of RAG in medicine to date. Eighteen medical experts contributed a total of 80,502 annotations, assessing 800 model outputs generated by GPT-4o and Llama-3.1-8B across 200 real-world patient and USMLE-style queries. We systematically decomposed the RAG pipeline into three components: (i) evidence retrieval (relevance of retrieved passages), (ii) evidence selection (accuracy of evidence usage), and (iii) response generation (factuality and completeness of outputs). Contrary to expectation, standard RAG often degraded performance: only 22% of top-16 passages were relevant, evidence selection remained weak (precision 41-43%, recall 27-49%), and factuality and completeness dropped by up to 6% and 5%, respectively, compared with non-RAG variants. Retrieval and evidence selection remain key failure points for the model, contributing to the overall performance drop. We further show that simple yet effective strategies, including evidence filtering and query reformulation, substantially mitigate these issues, improving performance on MedMCQA and MedXpertQA by up to 12% and 8.2%, respectively. These findings call for re-examining RAG's role in medicine and highlight the importance of stage-aware evaluation and deliberate system design for reliable medical LLM applications.

Kuangxiao Gu, Yuqian Zhou, Thomas Huang

Talking face synthesis has been widely studied in either appearance-based or warping-based methods. Previous works mostly utilize single face image as a source, and generate novel facial animations by merging other person's facial features. However, some facial regions like eyes or teeth, which may be hidden in the source image, can not be synthesized faithfully and stably. In this paper, We present a landmark driven two-stream network to generate faithful talking facial animation, in which more facial details are created, preserved and transferred from multiple source images instead of a single one. Specifically, we propose a network consisting of a learning and fetching stream. The fetching sub-net directly learns to attentively warp and merge facial regions from five source images of distinctive landmarks, while the learning pipeline renders facial organs from the training face space to compensate. Compared to baseline algorithms, extensive experiments demonstrate that the proposed method achieves a higher performance both quantitatively and qualitatively. Codes are at https://github.com/kgu3/FLNet_AAAI2020.

Yuqian Zhou, Jianbo Jiao, Haibin Huang et al.

Discriminative learning-based image denoisers have achieved promising performance on synthetic noises such as Additive White Gaussian Noise (AWGN). The synthetic noises adopted in most previous work are pixel-independent, but real noises are mostly spatially/channel-correlated and spatially/channel-variant. This domain gap yields unsatisfied performance on images with real noises if the model is only trained with AWGN. In this paper, we propose a novel approach to boost the performance of a real image denoiser which is trained only with synthetic pixel-independent noise data dominated by AWGN. First, we train a deep model that consists of a noise estimator and a denoiser with mixed AWGN and Random Value Impulse Noise (RVIN). We then investigate Pixel-shuffle Down-sampling (PD) strategy to adapt the trained model to real noises. Extensive experiments demonstrate the effectiveness and generalization of the proposed approach. Notably, our method achieves state-of-the-art performance on real sRGB images in the DND benchmark among models trained with synthetic noises. Codes are available at https://github.com/yzhouas/PD-Denoising-pytorch.

Jiahui Yu, Thomas Huang

We study how to set channel numbers in a neural network to achieve better accuracy under constrained resources (e.g., FLOPs, latency, memory footprint or model size). A simple and one-shot solution, named AutoSlim, is presented. Instead of training many network samples and searching with reinforcement learning, we train a single slimmable network to approximate the network accuracy of different channel configurations. We then iteratively evaluate the trained slimmable model and greedily slim the layer with minimal accuracy drop. By this single pass, we can obtain the optimized channel configurations under different resource constraints. We present experiments with MobileNet v1, MobileNet v2, ResNet-50 and RL-searched MNasNet on ImageNet classification. We show significant improvements over their default channel configurations. We also achieve better accuracy than recent channel pruning methods and neural architecture search methods. Notably, by setting optimized channel numbers, our AutoSlim-MobileNet-v2 at 305M FLOPs achieves 74.2% top-1 accuracy, 2.4% better than default MobileNet-v2 (301M FLOPs), and even 0.2% better than RL-searched MNasNet (317M FLOPs). Our AutoSlim-ResNet-50 at 570M FLOPs, without depthwise convolutions, achieves 1.3% better accuracy than MobileNet-v1 (569M FLOPs). Code and models will be available at: https://github.com/JiahuiYu/slimmable_networks

Jiahui Yu, Thomas Huang

Slimmable networks are a family of neural networks that can instantly adjust the runtime width. The width can be chosen from a predefined widths set to adaptively optimize accuracy-efficiency trade-offs at runtime. In this work, we propose a systematic approach to train universally slimmable networks (US-Nets), extending slimmable networks to execute at arbitrary width, and generalizing to networks both with and without batch normalization layers. We further propose two improved training techniques for US-Nets, named the sandwich rule and inplace distillation, to enhance training process and boost testing accuracy. We show improved performance of universally slimmable MobileNet v1 and MobileNet v2 on ImageNet classification task, compared with individually trained ones and 4-switch slimmable network baselines. We also evaluate the proposed US-Nets and improved training techniques on tasks of image super-resolution and deep reinforcement learning. Extensive ablation experiments on these representative tasks demonstrate the effectiveness of our proposed methods. Our discovery opens up the possibility to directly evaluate FLOPs-Accuracy spectrum of network architectures. Code and models are available at: https://github.com/JiahuiYu/slimmable_networks

Jiahui Yu, Linjie Yang, Ning Xu et al.

We present a simple and general method to train a single neural network executable at different widths (number of channels in a layer), permitting instant and adaptive accuracy-efficiency trade-offs at runtime. Instead of training individual networks with different width configurations, we train a shared network with switchable batch normalization. At runtime, the network can adjust its width on the fly according to on-device benchmarks and resource constraints, rather than downloading and offloading different models. Our trained networks, named slimmable neural networks, achieve similar (and in many cases better) ImageNet classification accuracy than individually trained models of MobileNet v1, MobileNet v2, ShuffleNet and ResNet-50 at different widths respectively. We also demonstrate better performance of slimmable models compared with individual ones across a wide range of applications including COCO bounding-box object detection, instance segmentation and person keypoint detection without tuning hyper-parameters. Lastly we visualize and discuss the learned features of slimmable networks. Code and models are available at: https://github.com/JiahuiYu/slimmable_networks

Yang Fu, Yunchao Wei, Guanshuo Wang et al.

Domain adaptation in person re-identification (re-ID) has always been a challenging task. In this work, we explore how to harness the natural similar characteristics existing in the samples from the target domain for learning to conduct person re-ID in an unsupervised manner. Concretely, we propose a Self-similarity Grouping (SSG) approach, which exploits the potential similarity (from global body to local parts) of unlabeled samples to automatically build multiple clusters from different views. These independent clusters are then assigned with labels, which serve as the pseudo identities to supervise the training process. We repeatedly and alternatively conduct such a grouping and training process until the model is stable. Despite the apparent simplify, our SSG outperforms the state-of-the-arts by more than 4.6% (DukeMTMC to Market1501) and 4.4% (Market1501 to DukeMTMC) in mAP, respectively. Upon our SSG, we further introduce a clustering-guided semisupervised approach named SSG ++ to conduct the one-shot domain adaption in an open set setting (i.e. the number of independent identities from the target domain is unknown). Without spending much effort on labeling, our SSG ++ can further promote the mAP upon SSG by 10.7% and 6.9%, respectively. Our Code is available at: https://github.com/OasisYang/SSG .

Bowen Cheng, Yunchao Wei, Rogerio Feris et al.

In this paper, we analyze failure cases of state-of-the-art detectors and observe that most hard false positives result from classification instead of localization and they have a large negative impact on the performance of object detectors. We conjecture there are three factors: (1) Shared feature representation is not optimal due to the mismatched goals of feature learning for classification and localization; (2) multi-task learning helps, yet optimization of the multi-task loss may result in sub-optimal for individual tasks; (3) large receptive field for different scales leads to redundant context information for small objects. We demonstrate the potential of detector classification power by a simple, effective, and widely-applicable Decoupled Classification Refinement (DCR) network. In particular, DCR places a separate classification network in parallel with the localization network (base detector). With ROI Pooling placed on the early stage of the classification network, we enforce an adaptive receptive field in DCR. During training, DCR samples hard false positives from the base detector and trains a strong classifier to refine classification results. During testing, DCR refines all boxes from the base detector. Experiments show competitive results on PASCAL VOC and COCO without any bells and whistles. Our codes are available at: https://github.com/bowenc0221/Decoupled-Classification-Refinement.

Tao Ruan, Ting Liu, Zilong Huang et al.

Human parsing has received considerable interest due to its wide application potentials. Nevertheless, it is still unclear how to develop an accurate human parsing system in an efficient and elegant way. In this paper, we identify several useful properties, including feature resolution, global context information and edge details, and perform rigorous analyses to reveal how to leverage them to benefit the human parsing task. The advantages of these useful properties finally result in a simple yet effective Context Embedding with Edge Perceiving (CE2P) framework for single human parsing. Our CE2P is end-to-end trainable and can be easily adopted for conducting multiple human parsing. Benefiting the superiority of CE2P, we achieved the 1st places on all three human parsing benchmarks. Without any bells and whistles, we achieved 56.50\% (mIoU), 45.31\% (mean $AP^r$) and 33.34\% ($AP^p_{0.5}$) in LIP, CIHP and MHP v2.0, which outperform the state-of-the-arts more than 2.06\%, 3.81\% and 1.87\%, respectively. We hope our CE2P will serve as a solid baseline and help ease future research in single/multiple human parsing. Code has been made available at \url{https://github.com/liutinglt/CE2P}.

Jiahui Yu, Yuchen Fan, Jianchao Yang et al.

In this report we demonstrate that with same parameters and computational budgets, models with wider features before ReLU activation have significantly better performance for single image super-resolution (SISR). The resulted SR residual network has a slim identity mapping pathway with wider (\(2\times\) to \(4\times\)) channels before activation in each residual block. To further widen activation (\(6\times\) to \(9\times\)) without computational overhead, we introduce linear low-rank convolution into SR networks and achieve even better accuracy-efficiency tradeoffs. In addition, compared with batch normalization or no normalization, we find training with weight normalization leads to better accuracy for deep super-resolution networks. Our proposed SR network \textit{WDSR} achieves better results on large-scale DIV2K image super-resolution benchmark in terms of PSNR with same or lower computational complexity. Based on WDSR, our method also won 1st places in NTIRE 2018 Challenge on Single Image Super-Resolution in all three realistic tracks. Experiments and ablation studies support the importance of wide activation for image super-resolution. Code is released at: https://github.com/JiahuiYu/wdsr_ntire2018

Jiahui Yu, Zhe Lin, Jimei Yang et al.

We present a generative image inpainting system to complete images with free-form mask and guidance. The system is based on gated convolutions learned from millions of images without additional labelling efforts. The proposed gated convolution solves the issue of vanilla convolution that treats all input pixels as valid ones, generalizes partial convolution by providing a learnable dynamic feature selection mechanism for each channel at each spatial location across all layers. Moreover, as free-form masks may appear anywhere in images with any shape, global and local GANs designed for a single rectangular mask are not applicable. Thus, we also present a patch-based GAN loss, named SN-PatchGAN, by applying spectral-normalized discriminator on dense image patches. SN-PatchGAN is simple in formulation, fast and stable in training. Results on automatic image inpainting and user-guided extension demonstrate that our system generates higher-quality and more flexible results than previous methods. Our system helps user quickly remove distracting objects, modify image layouts, clear watermarks and edit faces. Code, demo and models are available at: https://github.com/JiahuiYu/generative_inpainting

Yuqian Zhou, Kuangxiao Gu, Thomas Huang

A good representation for arbitrarily complicated data should have the capability of semantic generation, clustering and reconstruction. Previous research has already achieved impressive performance on either one. This paper aims at learning a disentangled representation effective for all of them in an unsupervised way. To achieve all the three tasks together, we learn the forward and inverse mapping between data and representation on the basis of a symmetric adversarial process. In theory, we minimize the upper bound of the two conditional entropy loss between the latent variables and the observations together to achieve the cycle consistency. The newly proposed RepGAN is tested on MNIST, fashionMNIST, CelebA, and SVHN datasets to perform unsupervised classification, generation and reconstruction tasks. The result demonstrates that RepGAN is able to learn a useful and competitive representation. To the author's knowledge, our work is the first one to achieve both a high unsupervised classification accuracy and low reconstruction error on MNIST. Codes are available at https://github.com/yzhouas/RepGAN-tensorflow.

Yang Fu, Yunchao Wei, Yuqian Zhou et al.

Despite the remarkable recent progress, person re-identification (Re-ID) approaches are still suffering from the failure cases where the discriminative body parts are missing. To mitigate such cases, we propose a simple yet effective Horizontal Pyramid Matching (HPM) approach to fully exploit various partial information of a given person, so that correct person candidates can be still identified even even some key parts are missing. Within the HPM, we make the following contributions to produce a more robust feature representation for the Re-ID task: 1) we learn to classify using partial feature representations at different horizontal pyramid scales, which successfully enhance the discriminative capabilities of various person parts; 2) we exploit average and max pooling strategies to account for person-specific discriminative information in a global-local manner. To validate the effectiveness of the proposed HPM, extensive experiments are conducted on three popular benchmarks, including Market-1501, DukeMTMC-ReID and CUHK03. In particular, we achieve mAP scores of 83.1%, 74.5% and 59.7% on these benchmarks, which are the new state-of-the-arts. Our code is available on Github

Zhiqiang Shen, Honghui Shi, Jiahui Yu et al.

In this paper, we present a simple and parameter-efficient drop-in module for one-stage object detectors like SSD when learning from scratch (i.e., without pre-trained models). We call our module GFR (Gated Feature Reuse), which exhibits two main advantages. First, we introduce a novel gate-controlled prediction strategy enabled by Squeeze-and-Excitation to adaptively enhance or attenuate supervision at different scales based on the input object size. As a result, our model is more effective in detecting diverse sizes of objects. Second, we propose a feature-pyramids structure to squeeze rich spatial and semantic features into a single prediction layer, which strengthens feature representation and reduces the number of parameters to learn. We apply the proposed structure on DSOD and SSD detection frameworks, and evaluate the performance on PASCAL VOC 2007, 2012 and COCO datasets. With fewer model parameters, GFR-DSOD outperforms the baseline DSOD by 1.4%, 1.1%, 1.7% and 0.6%, respectively. GFR-SSD also outperforms the original SSD and SSD with dense prediction by 3.6% and 2.8% on VOC 2007 dataset. Code is available at: https://github.com/szq0214/GFR-DSOD .

Qiao Jin, Zhizheng Wang, Yifan Yang et al.

Clinical calculators play a vital role in healthcare by offering accurate evidence-based predictions for various purposes such as prognosis. Nevertheless, their widespread utilization is frequently hindered by usability challenges, poor dissemination, and restricted functionality. Augmenting large language models with extensive collections of clinical calculators presents an opportunity to overcome these obstacles and improve workflow efficiency, but the scalability of the manual curation process poses a significant challenge. In response, we introduce AgentMD, a novel language agent capable of curating and applying clinical calculators across various clinical contexts. Using the published literature, AgentMD has automatically curated a collection of 2,164 diverse clinical calculators with executable functions and structured documentation, collectively named RiskCalcs. Manual evaluations show that RiskCalcs tools achieve an accuracy of over 80% on three quality metrics. At inference time, AgentMD can automatically select and apply the relevant RiskCalcs tools given any patient description. On the newly established RiskQA benchmark, AgentMD significantly outperforms chain-of-thought prompting with GPT-4 (87.7% vs. 40.9% in accuracy). Additionally, we also applied AgentMD to real-world clinical notes for analyzing both population-level and risk-level patient characteristics. In summary, our study illustrates the utility of language agents augmented with clinical calculators for healthcare analytics and patient care.

Wonkwang Lee, Whie Jung, Han Zhang et al.

Learning to predict the long-term future of video frames is notoriously challenging due to inherent ambiguities in the distant future and dramatic amplifications of prediction error through time. Despite the recent advances in the literature, existing approaches are limited to moderately short-term prediction (less than a few seconds), while extrapolating it to a longer future quickly leads to destruction in structure and content. In this work, we revisit hierarchical models in video prediction. Our method predicts future frames by first estimating a sequence of semantic structures and subsequently translating the structures to pixels by video-to-video translation. Despite the simplicity, we show that modeling structures and their dynamics in the discrete semantic structure space with a stochastic recurrent estimator leads to surprisingly successful long-term prediction. We evaluate our method on three challenging datasets involving car driving and human dancing, and demonstrate that it can generate complicated scene structures and motions over a very long time horizon (i.e., thousands frames), setting a new standard of video prediction with orders of magnitude longer prediction time than existing approaches. Full videos and codes are available at https://1konny.github.io/HVP/.

Md Yousuf Harun, Thomas Huang, Aaron T. Ohta

Embryo quality assessment based on morphological attributes is important for achieving higher pregnancy rates from in vitro fertilization (IVF). The accurate segmentation of the embryo's inner cell mass (ICM) and trophectoderm epithelium (TE) is important, as these parameters can help to predict the embryo viability and live birth potential. However, segmentation of the ICM and TE is difficult due to variations in their shape and similarities in their textures, both with each other and with their surroundings. To tackle this problem, a deep neural network (DNN) based segmentation approach was implemented. The DNN can identify the ICM region with 99.1% accuracy, 94.9% precision, 93.8% recall, a 94.3% Dice Coefficient, and a 89.3% Jaccard Index. It can extract the TE region with 98.3% accuracy, 91.8% precision, 93.2% recall, a 92.5% Dice Coefficient, and a 85.3% Jaccard Index.

Md Yousuf Harun, M Arifur Rahman, Joshua Mellinger et al.

Automating human preimplantation embryo grading offers the potential for higher success rates with in vitro fertilization (IVF) by providing new quantitative and objective measures of embryo quality. Current IVF procedures typically use only qualitative manual grading, which is limited in the identification of genetically abnormal embryos. The automatic quantitative assessment of blastocyst expansion can potentially improve sustained pregnancy rates and reduce health risks from abnormal pregnancies through a more accurate identification of genetic abnormality. The expansion rate of a blastocyst is an important morphological feature to determine the quality of a developing embryo. In this work, a deep learning based human blastocyst image segmentation method is presented, with the goal of facilitating the challenging task of segmenting irregularly shaped blastocysts. The type of blastocysts evaluated here has undergone laser ablation of the zona pellucida, which is required prior to trophectoderm biopsy. This complicates the manual measurements of the expanded blastocyst's size, which shows a correlation with genetic abnormalities. The experimental results on the test set demonstrate segmentation greatly improves the accuracy of expansion measurements, resulting in up to 99.4% accuracy, 98.1% precision, 98.8% recall, a 98.4% Dice Coefficient, and a 96.9% Jaccard Index.

Abdelrahman Abdelhamed, Mahmoud Afifi, Radu Timofte et al.

This paper reviews the NTIRE 2020 challenge on real image denoising with focus on the newly introduced dataset, the proposed methods and their results. The challenge is a new version of the previous NTIRE 2019 challenge on real image denoising that was based on the SIDD benchmark. This challenge is based on a newly collected validation and testing image datasets, and hence, named SIDD+. This challenge has two tracks for quantitatively evaluating image denoising performance in (1) the Bayer-pattern rawRGB and (2) the standard RGB (sRGB) color spaces. Each track ~250 registered participants. A total of 22 teams, proposing 24 methods, competed in the final phase of the challenge. The proposed methods by the participating teams represent the current state-of-the-art performance in image denoising targeting real noisy images. The newly collected SIDD+ datasets are publicly available at: https://bit.ly/siddplus_data.

Jiahui Yu, Pengchong Jin, Hanxiao Liu et al.

Neural architecture search (NAS) has shown promising results discovering models that are both accurate and fast. For NAS, training a one-shot model has become a popular strategy to rank the relative quality of different architectures (child models) using a single set of shared weights. However, while one-shot model weights can effectively rank different network architectures, the absolute accuracies from these shared weights are typically far below those obtained from stand-alone training. To compensate, existing methods assume that the weights must be retrained, finetuned, or otherwise post-processed after the search is completed. These steps significantly increase the compute requirements and complexity of the architecture search and model deployment. In this work, we propose BigNAS, an approach that challenges the conventional wisdom that post-processing of the weights is necessary to get good prediction accuracies. Without extra retraining or post-processing steps, we are able to train a single set of shared weights on ImageNet and use these weights to obtain child models whose sizes range from 200 to 1000 MFLOPs. Our discovered model family, BigNASModels, achieve top-1 accuracies ranging from 76.5% to 80.9%, surpassing state-of-the-art models in this range including EfficientNets and Once-for-All networks without extra retraining or post-processing. We present ablative study and analysis to further understand the proposed BigNASModels.

Hanchao Yu, Shanhui Sun, Haichao Yu et al.

Motion estimation of cardiac MRI videos is crucial for the evaluation of human heart anatomy and function. Recent researches show promising results with deep learning-based methods. In clinical deployment, however, they suffer dramatic performance drops due to mismatched distributions between training and testing datasets, commonly encountered in the clinical environment. On the other hand, it is arguably impossible to collect all representative datasets and to train a universal tracker before deployment. In this context, we proposed a novel fast online adaptive learning (FOAL) framework: an online gradient descent based optimizer that is optimized by a meta-learner. The meta-learner enables the online optimizer to perform a fast and robust adaptation. We evaluated our method through extensive experiments on two public clinical datasets. The results showed the superior performance of FOAL in accuracy compared to the offline-trained tracking method. On average, the FOAL took only $0.4$ second per video for online optimization.

Xiaofan Zhang, Haoming Lu, Cong Hao et al.

Object detection and tracking are challenging tasks for resource-constrained embedded systems. While these tasks are among the most compute-intensive tasks from the artificial intelligence domain, they are only allowed to use limited computation and memory resources on embedded devices. In the meanwhile, such resource-constrained implementations are often required to satisfy additional demanding requirements such as real-time response, high-throughput performance, and reliable inference accuracy. To overcome these challenges, we propose SkyNet, a hardware-efficient neural network to deliver the state-of-the-art detection accuracy and speed for embedded systems. Instead of following the common top-down flow for compact DNN (Deep Neural Network) design, SkyNet provides a bottom-up DNN design approach with comprehensive understanding of the hardware constraints at the very beginning to deliver hardware-efficient DNNs. The effectiveness of SkyNet is demonstrated by winning the competitive System Design Contest for low power object detection in the 56th IEEE/ACM Design Automation Conference (DAC-SDC), where our SkyNet significantly outperforms all other 100+ competitors: it delivers 0.731 Intersection over Union (IoU) and 67.33 frames per second (FPS) on a TX2 embedded GPU; and 0.716 IoU and 25.05 FPS on an Ultra96 embedded FPGA. The evaluation of SkyNet is also extended to GOT-10K, a recent large-scale high-diversity benchmark for generic object tracking in the wild. For state-of-the-art object trackers SiamRPN++ and SiamMask, where ResNet-50 is employed as the backbone, implementations using our SkyNet as the backbone DNN are 1.60X and 1.73X faster with better or similar accuracy when running on a 1080Ti GPU, and 37.20X smaller in terms of parameter size for significantly better memory and storage footprint.

Bowen Cheng, Liang-Chieh Chen, Yunchao Wei et al.

Multi-scale context module and single-stage encoder-decoder structure are commonly employed for semantic segmentation. The multi-scale context module refers to the operations to aggregate feature responses from a large spatial extent, while the single-stage encoder-decoder structure encodes the high-level semantic information in the encoder path and recovers the boundary information in the decoder path. In contrast, multi-stage encoder-decoder networks have been widely used in human pose estimation and show superior performance than their single-stage counterpart. However, few efforts have been attempted to bring this effective design to semantic segmentation. In this work, we propose a Semantic Prediction Guidance (SPG) module which learns to re-weight the local features through the guidance from pixel-wise semantic prediction. We find that by carefully re-weighting features across stages, a two-stage encoder-decoder network coupled with our proposed SPG module can significantly outperform its one-stage counterpart with similar parameters and computations. Finally, we report experimental results on the semantic segmentation benchmark Cityscapes, in which our SPGNet attains 81.1% on the test set using only 'fine' annotations.

Xiaofan Zhang, Cong Hao, Haoming Lu et al.

Developing artificial intelligence (AI) at the edge is always challenging, since edge devices have limited computation capability and memory resources but need to meet demanding requirements, such as real-time processing, high throughput performance, and high inference accuracy. To overcome these challenges, we propose SkyNet, an extremely lightweight DNN with 12 convolutional (Conv) layers and only 1.82 megabyte (MB) of parameters following a bottom-up DNN design approach. SkyNet is demonstrated in the 56th IEEE/ACM Design Automation Conference System Design Contest (DAC-SDC), a low power object detection challenge in images captured by unmanned aerial vehicles (UAVs). SkyNet won the first place award for both the GPU and FPGA tracks of the contest: we deliver 0.731 Intersection over Union (IoU) and 67.33 frames per second (FPS) on a TX2 GPU and deliver 0.716 IoU and 25.05 FPS on an Ultra96 FPGA.

Bowen Cheng, Rong Xiao, Jianfeng Wang et al.

We present a novel high frequency residual learning framework, which leads to a highly efficient multi-scale network (MSNet) architecture for mobile and embedded vision problems. The architecture utilizes two networks: a low resolution network to efficiently approximate low frequency components and a high resolution network to learn high frequency residuals by reusing the upsampled low resolution features. With a classifier calibration module, MSNet can dynamically allocate computation resources during inference to achieve a better speed and accuracy trade-off. We evaluate our methods on the challenging ImageNet-1k dataset and observe consistent improvements over different base networks. On ResNet-18 and MobileNet with alpha=1.0, MSNet gains 1.5% accuracy over both architectures without increasing computations. On the more efficient MobileNet with alpha=0.25, our method gains 3.8% accuracy with the same amount of computations.

Zhiqiang Shen, Mingyang Huang, Jianping Shi et al.

Unpaired Image-to-image Translation is a new rising and challenging vision problem that aims to learn a mapping between unaligned image pairs in diverse domains. Recent advances in this field like MUNIT and DRIT mainly focus on disentangling content and style/attribute from a given image first, then directly adopting the global style to guide the model to synthesize new domain images. However, this kind of approaches severely incurs contradiction if the target domain images are content-rich with multiple discrepant objects. In this paper, we present a simple yet effective instance-aware image-to-image translation approach (INIT), which employs the fine-grained local (instance) and global styles to the target image spatially. The proposed INIT exhibits three import advantages: (1) the instance-level objective loss can help learn a more accurate reconstruction and incorporate diverse attributes of objects; (2) the styles used for target domain of local/global areas are from corresponding spatial regions in source domain, which intuitively is a more reasonable mapping; (3) the joint training process can benefit both fine and coarse granularity and incorporates instance information to improve the quality of global translation. We also collect a large-scale benchmark for the new instance-level translation task. We observe that our synthetic images can even benefit real-world vision tasks like generic object detection.

Yang Fu, Xiaoyang Wang, Yunchao Wei et al.

In this work, we propose a novel Spatial-Temporal Attention (STA) approach to tackle the large-scale person re-identification task in videos. Different from the most existing methods, which simply compute representations of video clips using frame-level aggregation (e.g. average pooling), the proposed STA adopts a more effective way for producing robust clip-level feature representation. Concretely, our STA fully exploits those discriminative parts of one target person in both spatial and temporal dimensions, which results in a 2-D attention score matrix via inter-frame regularization to measure the importances of spatial parts across different frames. Thus, a more robust clip-level feature representation can be generated according to a weighted sum operation guided by the mined 2-D attention score matrix. In this way, the challenging cases for video-based person re-identification such as pose variation and partial occlusion can be well tackled by the STA. We conduct extensive experiments on two large-scale benchmarks, i.e. MARS and DukeMTMC-VideoReID. In particular, the mAP reaches 87.7% on MARS, which significantly outperforms the state-of-the-arts with a large margin of more than 11.6%.

Rui Qian, Yunchao Wei, Honghui Shi et al.

Semantic scene parsing is suffering from the fact that pixel-level annotations are hard to be collected. To tackle this issue, we propose a Point-based Distance Metric Learning (PDML) in this paper. PDML does not require dense annotated masks and only leverages several labeled points that are much easier to obtain to guide the training process. Concretely, we leverage semantic relationship among the annotated points by encouraging the feature representations of the intra- and inter-category points to keep consistent, i.e. points within the same category should have more similar feature representations compared to those from different categories. We formulate such a characteristic into a simple distance metric loss, which collaborates with the point-wise cross-entropy loss to optimize the deep neural networks. Furthermore, to fully exploit the limited annotations, distance metric learning is conducted across different training images instead of simply adopting an image-dependent manner. We conduct extensive experiments on two challenging scene parsing benchmarks of PASCAL-Context and ADE 20K to validate the effectiveness of our PDML, and competitive mIoU scores are achieved.

Xiaolin Zhang, Yunchao Wei, Yi Yang et al.

One-shot image semantic segmentation poses a challenging task of recognizing the object regions from unseen categories with only one annotated example as supervision. In this paper, we propose a simple yet effective Similarity Guidance network to tackle the One-shot (SG-One) segmentation problem. We aim at predicting the segmentation mask of a query image with the reference to one densely labeled support image of the same category. To obtain the robust representative feature of the support image, we firstly adopt a masked average pooling strategy for producing the guidance features by only taking the pixels belonging to the support image into account. We then leverage the cosine similarity to build the relationship between the guidance features and features of pixels from the query image. In this way, the possibilities embedded in the produced similarity maps can be adapted to guide the process of segmenting objects. Furthermore, our SG-One is a unified framework which can efficiently process both support and query images within one network and be learned in an end-to-end manner. We conduct extensive experiments on Pascal VOC 2012. In particular, our SGOne achieves the mIoU score of 46.3%, surpassing the baseline methods.

Ning Xu, Linjie Yang, Yuchen Fan et al.

Learning long-term spatial-temporal features are critical for many video analysis tasks. However, existing video segmentation methods predominantly rely on static image segmentation techniques, and methods capturing temporal dependency for segmentation have to depend on pretrained optical flow models, leading to suboptimal solutions for the problem. End-to-end sequential learning to explore spatialtemporal features for video segmentation is largely limited by the scale of available video segmentation datasets, i.e., even the largest video segmentation dataset only contains 90 short video clips. To solve this problem, we build a new large-scale video object segmentation dataset called YouTube Video Object Segmentation dataset (YouTube-VOS). Our dataset contains 4,453 YouTube video clips and 94 object categories. This is by far the largest video object segmentation dataset to our knowledge and has been released at http://youtube-vos.org. We further evaluate several existing state-of-the-art video object segmentation algorithms on this dataset which aims to establish baselines for the development of new algorithms in the future.

Ning Xu, Linjie Yang, Yuchen Fan et al.

Learning long-term spatial-temporal features are critical for many video analysis tasks. However, existing video segmentation methods predominantly rely on static image segmentation techniques, and methods capturing temporal dependency for segmentation have to depend on pretrained optical flow models, leading to suboptimal solutions for the problem. End-to-end sequential learning to explore spatial-temporal features for video segmentation is largely limited by the scale of available video segmentation datasets, i.e., even the largest video segmentation dataset only contains 90 short video clips. To solve this problem, we build a new large-scale video object segmentation dataset called YouTube Video Object Segmentation dataset (YouTube-VOS). Our dataset contains 3,252 YouTube video clips and 78 categories including common objects and human activities. This is by far the largest video object segmentation dataset to our knowledge and we have released it at https://youtube-vos.org. Based on this dataset, we propose a novel sequence-to-sequence network to fully exploit long-term spatial-temporal information in videos for segmentation. We demonstrate that our method is able to achieve the best results on our YouTube-VOS test set and comparable results on DAVIS 2016 compared to the current state-of-the-art methods. Experiments show that the large scale dataset is indeed a key factor to the success of our model.

Seunghoon Hong, Xinchen Yan, Thomas Huang et al.

Understanding, reasoning, and manipulating semantic concepts of images have been a fundamental research problem for decades. Previous work mainly focused on direct manipulation on natural image manifold through color strokes, key-points, textures, and holes-to-fill. In this work, we present a novel hierarchical framework for semantic image manipulation. Key to our hierarchical framework is that we employ a structured semantic layout as our intermediate representation for manipulation. Initialized with coarse-level bounding boxes, our structure generator first creates pixel-wise semantic layout capturing the object shape, object-object interactions, and object-scene relations. Then our image generator fills in the pixel-level textures guided by the semantic layout. Such framework allows a user to manipulate images at object-level by adding, removing, and moving one bounding box at a time. Experimental evaluations demonstrate the advantages of the hierarchical manipulation framework over existing image generation and context hole-filing models, both qualitatively and quantitatively. Benefits of the hierarchical framework are further demonstrated in applications such as semantic object manipulation, interactive image editing, and data-driven image manipulation.

Xiaolin Zhang, Yunchao Wei, Guoliang Kang et al.

Weakly supervised methods usually generate localization results based on attention maps produced by classification networks. However, the attention maps exhibit the most discriminative parts of the object which are small and sparse. We propose to generate Self-produced Guidance (SPG) masks which separate the foreground, the object of interest, from the background to provide the classification networks with spatial correlation information of pixels. A stagewise approach is proposed to incorporate high confident object regions to learn the SPG masks. The high confident regions within attention maps are utilized to progressively learn the SPG masks. The masks are then used as an auxiliary pixel-level supervision to facilitate the training of classification networks. Extensive experiments on ILSVRC demonstrate that SPG is effective in producing high-quality object localizations maps. Particularly, the proposed SPG achieves the Top-1 localization error rate of 43.83% on the ILSVRC validation set, which is a new state-of-the-art error rate.

Yunchao Wei, Zhiqiang Shen, Bowen Cheng et al.

This work provides a simple approach to discover tight object bounding boxes with only image-level supervision, called Tight box mining with Surrounding Segmentation Context (TS2C). We observe that object candidates mined through current multiple instance learning methods are usually trapped to discriminative object parts, rather than the entire object. TS2C leverages surrounding segmentation context derived from weakly-supervised segmentation to suppress such low-quality distracting candidates and boost the high-quality ones. Specifically, TS2C is developed based on two key properties of desirable bounding boxes: 1) high purity, meaning most pixels in the box are with high object response, and 2) high completeness, meaning the box covers high object response pixels comprehensively. With such novel and computable criteria, more tight candidates can be discovered for learning a better object detector. With TS2C, we obtain 48.0% and 44.4% mAP scores on VOC 2007 and 2012 benchmarks, which are the new state-of-the-arts.

Yuqian Zhou, Ding Liu, Thomas Huang

Face detection is a well-explored problem. Many challenges on face detectors like extreme pose, illumination, low resolution and small scales are studied in the previous work. However, previous proposed models are mostly trained and tested on good-quality images which are not always the case for practical applications like surveillance systems. In this paper, we first review the current state-of-the-art face detectors and their performance on benchmark dataset FDDB, and compare the design protocols of the algorithms. Secondly, we investigate their performance degradation while testing on low-quality images with different levels of blur, noise, and contrast. Our results demonstrate that both hand-crafted and deep-learning based face detectors are not robust enough for low-quality images. It inspires researchers to produce more robust design for face detection in the wild.

Xiaolin Zhang, Yunchao Wei, Jiashi Feng et al.

In this work, we propose Adversarial Complementary Learning (ACoL) to automatically localize integral objects of semantic interest with weak supervision. We first mathematically prove that class localization maps can be obtained by directly selecting the class-specific feature maps of the last convolutional layer, which paves a simple way to identify object regions. We then present a simple network architecture including two parallel-classifiers for object localization. Specifically, we leverage one classification branch to dynamically localize some discriminative object regions during the forward pass. Although it is usually responsive to sparse parts of the target objects, this classifier can drive the counterpart classifier to discover new and complementary object regions by erasing its discovered regions from the feature maps. With such an adversarial learning, the two parallel-classifiers are forced to leverage complementary object regions for classification and can finally generate integral object localization together. The merits of ACoL are mainly two-fold: 1) it can be trained in an end-to-end manner; 2) dynamically erasing enables the counterpart classifier to discover complementary object regions more effectively. We demonstrate the superiority of our ACoL approach in a variety of experiments. In particular, the Top-1 localization error rate on the ILSVRC dataset is 45.14%, which is the new state-of-the-art.

Bowen Cheng, Yunchao Wei, Honghui Shi et al.

Recent region-based object detectors are usually built with separate classification and localization branches on top of shared feature extraction networks. In this paper, we analyze failure cases of state-of-the-art detectors and observe that most hard false positives result from classification instead of localization. We conjecture that: (1) Shared feature representation is not optimal due to the mismatched goals of feature learning for classification and localization; (2) multi-task learning helps, yet optimization of the multi-task loss may result in sub-optimal for individual tasks; (3) large receptive field for different scales leads to redundant context information for small objects.We demonstrate the potential of detector classification power by a simple, effective, and widely-applicable Decoupled Classification Refinement (DCR) network. DCR samples hard false positives from the base classifier in Faster RCNN and trains a RCNN-styled strong classifier. Experiments show new state-of-the-art results on PASCAL VOC and COCO without any bells and whistles.

Ning Xu, Brian Price, Scott Cohen et al.

Most previous bounding-box-based segmentation methods assume the bounding box tightly covers the object of interest. However it is common that a rectangle input could be too large or too small. In this paper, we propose a novel segmentation approach that uses a rectangle as a soft constraint by transforming it into an Euclidean distance map. A convolutional encoder-decoder network is trained end-to-end by concatenating images with these distance maps as inputs and predicting the object masks as outputs. Our approach gets accurate segmentation results given sloppy rectangles while being general for both interactive segmentation and instance segmentation. We show our network extends to curve-based input without retraining. We further apply our network to instance-level semantic segmentation and resolve any overlap using a conditional random field. Experiments on benchmark datasets demonstrate the effectiveness of the proposed approaches.

Guo-Jun Qi, Wei Liu, Charu Aggarwal et al.

In this paper, we present a label transfer model from texts to images for image classification tasks. The problem of image classification is often much more challenging than text classification. On one hand, labeled text data is more widely available than the labeled images for classification tasks. On the other hand, text data tends to have natural semantic interpretability, and they are often more directly related to class labels. On the contrary, the image features are not directly related to concepts inherent in class labels. One of our goals in this paper is to develop a model for revealing the functional relationships between text and image features as to directly transfer intermodal and intramodal labels to annotate the images. This is implemented by learning a transfer function as a bridge to propagate the labels between two multimodal spaces. However, the intermodal label transfers could be undermined by blindly transferring the labels of noisy texts to annotate images. To mitigate this problem, we present an intramodal label transfer process, which complements the intermodal label transfer by transferring the image labels instead when relevant text is absent from the source corpus. In addition, we generalize the inter-modal label transfer to zero-shot learning scenario where there are only text examples available to label unseen classes of images without any positive image examples. We evaluate our algorithm on an image classification task and show the effectiveness with respect to the other compared algorithms.

Ning Xu, Brian Price, Scott Cohen et al.

Image matting is a fundamental computer vision problem and has many applications. Previous algorithms have poor performance when an image has similar foreground and background colors or complicated textures. The main reasons are prior methods 1) only use low-level features and 2) lack high-level context. In this paper, we propose a novel deep learning based algorithm that can tackle both these problems. Our deep model has two parts. The first part is a deep convolutional encoder-decoder network that takes an image and the corresponding trimap as inputs and predict the alpha matte of the image. The second part is a small convolutional network that refines the alpha matte predictions of the first network to have more accurate alpha values and sharper edges. In addition, we also create a large-scale image matting dataset including 49300 training images and 1000 testing images. We evaluate our algorithm on the image matting benchmark, our testing set, and a wide variety of real images. Experimental results clearly demonstrate the superiority of our algorithm over previous methods.

Ding Liu, Zhaowen Wang, Nasser Nasrabadi et al.

Single image super-resolution (SR) is an ill-posed problem which aims to recover high-resolution (HR) images from their low-resolution (LR) observations. The crux of this problem lies in learning the complex mapping between low-resolution patches and the corresponding high-resolution patches. Prior arts have used either a mixture of simple regression models or a single non-linear neural network for this propose. This paper proposes the method of learning a mixture of SR inference modules in a unified framework to tackle this problem. Specifically, a number of SR inference modules specialized in different image local patterns are first independently applied on the LR image to obtain various HR estimates, and the resultant HR estimates are adaptively aggregated to form the final HR image. By selecting neural networks as the SR inference module, the whole procedure can be incorporated into a unified network and be optimized jointly. Extensive experiments are conducted to investigate the relation between restoration performance and different network architectures. Compared with other current image SR approaches, our proposed method achieves state-of-the-arts restoration results on a wide range of images consistently while allowing more flexible design choices. The source codes are available in http://www.ifp.illinois.edu/~dingliu2/accv2016.

Jiahui Yu, Yuning Jiang, Zhangyang Wang et al.

In present object detection systems, the deep convolutional neural networks (CNNs) are utilized to predict bounding boxes of object candidates, and have gained performance advantages over the traditional region proposal methods. However, existing deep CNN methods assume the object bounds to be four independent variables, which could be regressed by the $\ell_2$ loss separately. Such an oversimplified assumption is contrary to the well-received observation, that those variables are correlated, resulting to less accurate localization. To address the issue, we firstly introduce a novel Intersection over Union ($IoU$) loss function for bounding box prediction, which regresses the four bounds of a predicted box as a whole unit. By taking the advantages of $IoU$ loss and deep fully convolutional networks, the UnitBox is introduced, which performs accurate and efficient localization, shows robust to objects of varied shapes and scales, and converges fast. We apply UnitBox on face detection task and achieve the best performance among all published methods on the FDDB benchmark.

Xianming Liu, Min-Hsuan Tsai, Thomas Huang

With the incredibly growing amount of multimedia data shared on the social media platforms, recommender systems have become an important necessity to ease users' burden on the information overload. In such a scenario, extensive amount of heterogeneous information such as tags, image content, in addition to the user-to-item preferences, is extremely valuable for making effective recommendations. In this paper, we explore a novel hybrid algorithm termed {\em STM}, for image recommendation. STM jointly considers the problem of image content analysis with the users' preferences on the basis of sparse representation. STM is able to tackle the challenges of highly sparse user feedbacks and cold-start problmes in the social network scenario. In addition, our model is based on the classical probabilistic matrix factorization and can be easily extended to incorporate other useful information such as the social relationships. We evaluate our approach with a newly collected 0.3 million social image data set from Flickr. The experimental results demonstrate that sparse topic modeling of the image content leads to more effective recommendations, , with a significant performance gain over the state-of-the-art alternatives.

Ning Xu, Brian Price, Scott Cohen et al.

Interactive object selection is a very important research problem and has many applications. Previous algorithms require substantial user interactions to estimate the foreground and background distributions. In this paper, we present a novel deep learning based algorithm which has a much better understanding of objectness and thus can reduce user interactions to just a few clicks. Our algorithm transforms user provided positive and negative clicks into two Euclidean distance maps which are then concatenated with the RGB channels of images to compose (image, user interactions) pairs. We generate many of such pairs by combining several random sampling strategies to model user click patterns and use them to fine tune deep Fully Convolutional Networks (FCNs). Finally the output probability maps of our FCN 8s model is integrated with graph cut optimization to refine the boundary segments. Our model is trained on the PASCAL segmentation dataset and evaluated on other datasets with different object classes. Experimental results on both seen and unseen objects clearly demonstrate that our algorithm has a good generalization ability and is superior to all existing interactive object selection approaches.

Zhaowen Wang, Ding Liu, Jianchao Yang et al.

Deep learning techniques have been successfully applied in many areas of computer vision, including low-level image restoration problems. For image super-resolution, several models based on deep neural networks have been recently proposed and attained superior performance that overshadows all previous handcrafted models. The question then arises whether large-capacity and data-driven models have become the dominant solution to the ill-posed super-resolution problem. In this paper, we argue that domain expertise represented by the conventional sparse coding model is still valuable, and it can be combined with the key ingredients of deep learning to achieve further improved results. We show that a sparse coding model particularly designed for super-resolution can be incarnated as a neural network, and trained in a cascaded structure from end to end. The interpretation of the network based on sparse coding leads to much more efficient and effective training, as well as a reduced model size. Our model is evaluated on a wide range of images, and shows clear advantage over existing state-of-the-art methods in terms of both restoration accuracy and human subjective quality.

Zhaowen Wang, Jianchao Yang, Zhe Lin et al.

Classifying large-scale image data into object categories is an important problem that has received increasing research attention. Given the huge amount of data, non-parametric approaches such as nearest neighbor classifiers have shown promising results, especially when they are underpinned by a learned distance or similarity measurement. Although metric learning has been well studied in the past decades, most existing algorithms are impractical to handle large-scale data sets. In this paper, we present an image similarity learning method that can scale well in both the number of images and the dimensionality of image descriptors. To this end, similarity comparison is restricted to each sample's local neighbors and a discriminative similarity measure is induced from large margin neighborhood embedding. We also exploit the ensemble of projections so that high-dimensional features can be processed in a set of lower-dimensional subspaces in parallel without much performance compromise. The similarity function is learned online using a stochastic gradient descent algorithm in which the triplet sampling strategy is customized for quick convergence of classification performance. The effectiveness of our proposed model is validated on several data sets with scales varying from tens of thousands to one million images. Recognition accuracies competitive with the state-of-the-art performance are achieved with much higher efficiency and scalability.

Thomas Paine, Hailin Jin, Jianchao Yang et al.

The ability to train large-scale neural networks has resulted in state-of-the-art performance in many areas of computer vision. These results have largely come from computational break throughs of two forms: model parallelism, e.g. GPU accelerated training, which has seen quick adoption in computer vision circles, and data parallelism, e.g. A-SGD, whose large scale has been used mostly in industry. We report early experiments with a system that makes use of both model parallelism and data parallelism, we call GPU A-SGD. We show using GPU A-SGD it is possible to speed up training of large convolutional neural networks useful for computer vision. We believe GPU A-SGD will make it possible to train larger networks on larger training sets in a reasonable amount of time.

Guo-Jun Qi, Charu Aggarwal, Pierre Moulin et al.

Collective intelligence, which aggregates the shared information from large crowds, is often negatively impacted by unreliable information sources with the low quality data. This becomes a barrier to the effective use of collective intelligence in a variety of applications. In order to address this issue, we propose a probabilistic model to jointly assess the reliability of sources and find the true data. We observe that different sources are often not independent of each other. Instead, sources are prone to be mutually influenced, which makes them dependent when sharing information with each other. High dependency between sources makes collective intelligence vulnerable to the overuse of redundant (and possibly incorrect) information from the dependent sources. Thus, we reveal the latent group structure among dependent sources, and aggregate the information at the group level rather than from individual sources directly. This can prevent the collective intelligence from being inappropriately dominated by dependent sources. We will also explicitly reveal the reliability of groups, and minimize the negative impacts of unreliable groups. Experimental results on real-world data sets show the effectiveness of the proposed approach with respect to existing algorithms.

Zhaowen Wang, Jinjun Wang, Jing Xiao et al.

This paper introduces a probabilistic graphical model for continuous action recognition with two novel components: substructure transition model and discriminative boundary model. The first component encodes the sparse and global temporal transition prior between action primitives in state-space model to handle the large spatial-temporal variations within an action class. The second component enforces the action duration constraint in a discriminative way to locate the transition boundaries between actions more accurately. The two components are integrated into a unified graphical structure to enable effective training and inference. Our comprehensive experimental results on both public and in-house datasets show that, with the capability to incorporate additional information that had not been explicitly or efficiently modeled by previous methods, our proposed algorithm achieved significantly improved performance for continuous action recognition.