The study of 3D hyperspectral image (HSI) reconstruction refers to the inverse process of snapshot compressive imaging, during which the optical system, e.g., the coded aperture snapshot spectral imaging (CASSI) system, captures the 3D spatial-spectral signal and encodes it to a 2D measurement. While numerous sophisticated neural networks have been elaborated for end-to-end reconstruction, trade-offs still need to be made among performance, efficiency (training and inference time), and feasibility (the ability of restoring high resolution HSI on limited GPU memory). This raises a challenge to design a new baseline to conjointly meet the above requirements. In this paper, we fill in this blank by proposing a Spatial/Spectral Invariant Residual U-Net, namely SSI-ResU-Net. It differentiates with U-Net in three folds--1) scale/spectral-invariant learning, 2) nested residual learning, and 3) computational efficiency. Benefiting from these three modules, the proposed SSI-ResU-Net outperforms the current state-of-the-art method TSA-Net by over 3 dB in PSNR and 0.036 in SSIM while only using 2.82% trainable parameters. To the greatest extent, SSI-ResU-Net achieves competing performance with over 77.3% reduction in terms of floating-point operations (FLOPs), which for the first time, makes high-resolution HSI reconstruction feasible under practical application scenarios. Code and pre-trained models are made available at https://github.com/Jiamian-Wang/HSI_baseline.
Multimodal target/aspect sentiment classification combines multimodal sentiment analysis and aspect/target sentiment classification. The goal of the task is to combine vision and language to understand the sentiment towards a target entity in a sentence. Twitter is an ideal setting for the task because it is inherently multimodal, highly emotional, and affects real world events. However, multimodal tweets are short and accompanied by complex, possibly irrelevant images. We introduce a two-stream model that translates images in input space using an object-aware transformer followed by a single-pass non-autoregressive text generation approach. We then leverage the translation to construct an auxiliary sentence that provides multimodal information to a language model. Our approach increases the amount of text available to the language model and distills the object-level information in complex images. We achieve state-of-the-art performance on two multimodal Twitter datasets without modifying the internals of the language model to accept multimodal data, demonstrating the effectiveness of our translation. In addition, we explain a failure mode of a popular approach for aspect sentiment analysis when applied to tweets. Our code is available at \textcolor{blue}{\url{https://github.com/codezakh/exploiting-BERT-thru-translation}}.
Multimodal target/aspect sentiment classification combines multimodal sentiment analysis and aspect/target sentiment classification. The goal of the task is to combine vision and language to understand the sentiment towards a target entity in a sentence. Twitter is an ideal setting for the task because it is inherently multimodal, highly emotional, and affects real world events. However, multimodal tweets are short and accompanied by complex, possibly irrelevant images. We introduce a two-stream model that translates images in input space using an object-aware transformer followed by a single-pass non-autoregressive text generation approach. We then leverage the translation to construct an auxiliary sentence that provides multimodal information to a language model. Our approach increases the amount of text available to the language model and distills the object-level information in complex images. We achieve state-of-the-art performance on two multimodal Twitter datasets without modifying the internals of the language model to accept multimodal data, demonstrating the effectiveness of our translation. In addition, we explain a failure mode of a popular approach for aspect sentiment analysis when applied to tweets. Our code is available at \textcolor{blue}{\url{https://github.com/codezakh/exploiting-BERT-thru-translation}}.
Adaptive gradient methods, such as \textsc{Adam}, have achieved tremendous success in machine learning. Scaling gradients by square roots of the running averages of squared past gradients, such methods are able to attain rapid training of modern deep neural networks. Nevertheless, they are observed to generalize worse than stochastic gradient descent (\textsc{SGD}) and tend to be trapped in local minima at an early stage during training. Intriguingly, we discover that substituting the gradient in the preconditioner term with the momentumized version in \textsc{Adam} can well solve the issues. The intuition is that gradient with momentum contains more accurate directional information and therefore its second moment estimation is a better choice for scaling than raw gradient's. Thereby we propose \textsc{AdaMomentum} as a new optimizer reaching the goal of training faster while generalizing better. We further develop a theory to back up the improvement in optimization and generalization and provide convergence guarantee under both convex and nonconvex settings. Extensive experiments on various models and tasks demonstrate that \textsc{AdaMomentum} exhibits comparable performance to \textsc{SGD} on vision tasks, and achieves state-of-the-art results consistently on other tasks including language processing.
Several recent works [40, 24] observed an interesting phenomenon in neural network pruning: A larger finetuning learning rate can improve the final performance significantly. Unfortunately, the reason behind it remains elusive up to date. This paper is meant to explain it through the lens of dynamical isometry [42]. Specifically, we examine neural network pruning from an unusual perspective: pruning as initialization for finetuning, and ask whether the inherited weights serve as a good initialization for the finetuning? The insights from dynamical isometry suggest a negative answer. Despite its critical role, this issue has not been well-recognized by the community so far. In this paper, we will show the understanding of this problem is very important -- on top of explaining the aforementioned mystery about the larger finetuning rate, it also unveils the mystery about the value of pruning [5, 30]. Besides a clearer theoretical understanding of pruning, resolving the problem can also bring us considerable performance benefits in practice.
Domain adaptation enhances generalizability of a model across domains with domain shifts. Most research effort has been spent on Unsupervised Domain Adaption (UDA) which trains a model jointly with labeled source data and unlabeled target data. This paper studies how much it can help address domain shifts if we further have a few target samples (e.g., one sample per class) labeled. This is the so-called semi-supervised domain adaptation (SSDA) problem and the few labeled target samples are termed as ``landmarks''. To explore the full potential of landmarks, we incorporate a prototypical alignment (PA) module which calculates a target prototype for each class from the landmarks; source samples are then aligned with the target prototype from the same class. To further alleviate label scarcity, we propose a data augmentation based solution. Specifically, we severely perturb the labeled images, making PA non-trivial to achieve and thus promoting model generalizability. Moreover, we apply consistency learning on unlabeled target images, by perturbing each image with light transformations and strong transformations. Then, the strongly perturbed image can enjoy ``supervised-like'' training using the pseudo label inferred from the lightly perturbed one. Experiments show that the proposed method, though simple, reaches significant performance gains over state-of-the-art methods, and enjoys the flexibility of being able to serve as a plug-and-play component to various existing UDA methods and improve adaptation performance with landmarks provided. Our code is available at \url{https://github.com/kailigo/pacl}.
Cross-Domain Detection (XDD) aims to train an object detector using labeled image from a source domain but have good performance in the target domain with only unlabeled images. Existing approaches achieve this either by aligning the feature maps or the region proposals from the two domains, or by transferring the style of source images to that of target image. Contrasted with prior work, this paper provides a complementary solution to align domains by learning the same auxiliary tasks in both domains simultaneously. These auxiliary tasks push image from both domains towards shared spaces, which bridges the domain gap. Specifically, this paper proposes Rotation Prediction and Consistency Learning (PRCL), a framework complementing existing XDD methods for domain alignment by leveraging the two auxiliary tasks. The first one encourages the model to extract region proposals from foreground regions by rotating an image and predicting the rotation angle from the extracted region proposals. The second task encourages the model to be robust to changes in the image space by optimizing the model to make consistent class predictions for region proposals regardless of image perturbations. Experiments show the detection performance can be consistently and significantly enhanced by applying the two proposed tasks to existing XDD methods.
In this paper, we address the space-time video super-resolution, which aims at generating a high-resolution (HR) slow-motion video from a low-resolution (LR) and low frame rate (LFR) video sequence. A na\"ive method is to decompose it into two sub-tasks: video frame interpolation (VFI) and video super-resolution (VSR). Nevertheless, temporal interpolation and spatial upscaling are intra-related in this problem. Two-stage approaches cannot fully make use of this natural property. Besides, state-of-the-art VFI or VSR deep networks usually have a large frame reconstruction module in order to obtain high-quality photo-realistic video frames, which makes the two-stage approaches have large models and thus be relatively time-consuming. To overcome the issues, we present a one-stage space-time video super-resolution framework, which can directly reconstruct an HR slow-motion video sequence from an input LR and LFR video. Instead of reconstructing missing LR intermediate frames as VFI models do, we temporally interpolate LR frame features of the missing LR frames capturing local temporal contexts by a feature temporal interpolation module. Extensive experiments on widely used benchmarks demonstrate that the proposed framework not only achieves better qualitative and quantitative performance on both clean and noisy LR frames but also is several times faster than recent state-of-the-art two-stage networks. The source code is released in https://github.com/Mukosame/Zooming-Slow-Mo-CVPR-2020 .
Sign language is used by deaf or speech impaired people to communicate and requires great effort to master. Sign Language Recognition (SLR) aims to bridge between sign language users and others by recognizing words from given videos. It is an important yet challenging task since sign language is performed with fast and complex movement of hand gestures, body posture, and even facial expressions. Recently, skeleton-based action recognition attracts increasing attention due to the independence on subject and background variation. Furthermore, it can be a strong complement to RGB/D modalities to boost the overall recognition rate. However, skeleton-based SLR is still under exploration due to the lack of annotations on hand keypoints. Some efforts have been made to use hand detectors with pose estimators to extract hand key points and learn to recognize sign language via a Recurrent Neural Network, but none of them outperforms RGB-based methods. To this end, we propose a novel Skeleton Aware Multi-modal SLR framework (SAM-SLR) to further improve the recognition rate. Specifically, we propose a Sign Language Graph Convolution Network (SL-GCN) to model the embedded dynamics and propose a novel Separable Spatial-Temporal Convolution Network (SSTCN) to exploit skeleton features. Our skeleton-based method achieves a higher recognition rate compared with all other single modalities. Moreover, our proposed SAM-SLR framework can further enhance the performance by assembling our skeleton-based method with other RGB and depth modalities. As a result, SAM-SLR achieves the highest performance in both RGB (98.42%) and RGB-D (98.53%) tracks in 2021 Looking at People Large Scale Signer Independent Isolated SLR Challenge. Our code is available at https://github.com/jackyjsy/CVPR21Chal-SLR