Recently, self-supervised pretraining has achieved impressive results in end-to-end (E2E) automatic speech recognition (ASR). However, the dominant sequence-to-sequence (S2S) E2E model is still hard to fully utilize the self-supervised pre-training methods because its decoder is conditioned on acoustic representation thus cannot be pretrained separately. In this paper, we propose a pretrained Transformer (Preformer) S2S ASR architecture based on hybrid CTC/attention E2E models to fully utilize the pretrained acoustic models (AMs) and language models (LMs). In our framework, the encoder is initialized with a pretrained AM (wav2vec2.0). The Preformer leverages CTC as an auxiliary task during training and inference. Furthermore, we design a one-cross decoder (OCD), which relaxes the dependence on acoustic representations so that it can be initialized with pretrained LM (DistilGPT2). Experiments are conducted on the AISHELL-1 corpus and achieve a $4.6\%$ character error rate (CER) on the test set. Compared with our vanilla hybrid CTC/attention Transformer baseline, our proposed CTC/attention-based Preformer yields $27\%$ relative CER reduction. To the best of our knowledge, this is the first work to utilize both pretrained AM and LM in a S2S ASR system.
Existing studies on salient object detection (SOD) focus on extracting distinct objects with edge information and aggregating multi-level features to improve SOD performance. To achieve satisfactory performance, the methods employ refined edge information and low multi-level discrepancy. However, both performance gain and computational efficiency cannot be attained, which has motivated us to study the inefficiencies in existing encoder-decoder structures to avoid this trade-off. We propose TRACER, which detects salient objects with explicit edges by incorporating attention guided tracing modules. We employ a masked edge attention module at the end of the first encoder using a fast Fourier transform to propagate the refined edge information to the downstream feature extraction. In the multi-level aggregation phase, the union attention module identifies the complementary channel and important spatial information. To improve the decoder performance and computational efficiency, we minimize the decoder block usage with object attention module. This module extracts undetected objects and edge information from refined channels and spatial representations. Subsequently, we propose an adaptive pixel intensity loss function to deal with the relatively important pixels unlike conventional loss functions which treat all pixels equally. A comparison with 13 existing methods reveals that TRACER achieves state-of-the-art performance on five benchmark datasets. We have released TRACER at https://github.com/Karel911/TRACER.
Long story generation (LSG) is one of the coveted goals in natural language processing. Different from most text generation tasks, LSG requires to output a long story of rich content based on a much shorter text input, and often suffers from information sparsity. In this paper, we propose \emph{TopNet} to alleviate this problem, by leveraging the recent advances in neural topic modeling to obtain high-quality skeleton words to complement the short input. In particular, instead of directly generating a story, we first learn to map the short text input to a low-dimensional topic distribution (which is pre-assigned by a topic model). Based on this latent topic distribution, we can use the reconstruction decoder of the topic model to sample a sequence of inter-related words as a skeleton for the story. Experiments on two benchmark datasets show that our proposed framework is highly effective in skeleton word selection and significantly outperforms the state-of-the-art models in both automatic evaluation and human evaluation.
Few-shot classification aims to adapt classifiers to novel classes with a few training samples. However, the insufficiency of training data may cause a biased estimation of feature distribution in a certain class. To alleviate this problem, we present a simple yet effective feature rectification method by exploring the category correlation between novel and base classes as the prior knowledge. We explicitly capture such correlation by mapping features into a latent vector with dimension matching the number of base classes, treating it as the logarithm probability of the feature over base classes. Based on this latent vector, the rectified feature is directly constructed by a decoder, which we expect maintaining category-related information while removing other stochastic factors, and consequently being closer to its class centroid. Furthermore, by changing the temperature value in softmax, we can re-balance the feature rectification and reconstruction for better performance. Our method is generic, flexible and agnostic to any feature extractor and classifier, readily to be embedded into existing FSL approaches. Experiments verify that our method is capable of rectifying biased features, especially when the feature is far from the class centroid. The proposed approach consistently obtains considerable performance gains on three widely used benchmarks, evaluated with different backbones and classifiers. The code will be made public.
We propose a solution to the novel task of rendering sharp videos from new viewpoints from a single motion-blurred image of a face. Our method handles the complexity of face blur by implicitly learning the geometry and motion of faces through the joint training on three large datasets: FFHQ and 300VW, which are publicly available, and a new Bern Multi-View Face Dataset (BMFD) that we built. The first two datasets provide a large variety of faces and allow our model to generalize better. BMFD instead allows us to introduce multi-view constraints, which are crucial to synthesizing sharp videos from a new camera view. It consists of high frame rate synchronized videos from multiple views of several subjects displaying a wide range of facial expressions. We use the high frame rate videos to simulate realistic motion blur through averaging. Thanks to this dataset, we train a neural network to reconstruct a 3D video representation from a single image and the corresponding face gaze. We then provide a camera viewpoint relative to the estimated gaze and the blurry image as input to an encoder-decoder network to generate a video of sharp frames with a novel camera viewpoint. We demonstrate our approach on test subjects of our multi-view dataset and VIDTIMIT.
BERT-type structure has led to the revolution of vision-language pre-training and the achievement of state-of-the-art results on numerous vision-language downstream tasks. Existing solutions dominantly capitalize on the multi-modal inputs with mask tokens to trigger mask-based proxy pre-training tasks (e.g., masked language modeling and masked object/frame prediction). In this work, we argue that such masked inputs would inevitably introduce noise for cross-modal matching proxy task, and thus leave the inherent vision-language association under-explored. As an alternative, we derive a particular form of cross-modal proxy objective for video-language pre-training, i.e., Contrastive Cross-modal matching and denoising (CoCo). By viewing the masked frame/word sequences as the noisy augmentation of primary unmasked ones, CoCo strengthens video-language association by simultaneously pursuing inter-modal matching and intra-modal denoising between masked and unmasked inputs in a contrastive manner. Our CoCo proxy objective can be further integrated into any BERT-type encoder-decoder structure for video-language pre-training, named as Contrastive Cross-modal BERT (CoCo-BERT). We pre-train CoCo-BERT on TV dataset and a newly collected large-scale GIF video dataset (ACTION). Through extensive experiments over a wide range of downstream tasks (e.g., cross-modal retrieval, video question answering, and video captioning), we demonstrate the superiority of CoCo-BERT as a pre-trained structure.
Keyphrase generation aims at generating important phrases (keyphrases) that best describe a given document. In scholarly domains, current approaches have largely used only the title and abstract of the articles to generate keyphrases. In this paper, we comprehensively explore whether the integration of additional information from the full text of a given article or from semantically similar articles can be helpful for a neural keyphrase generation model or not. We discover that adding sentences from the full text, particularly in the form of the extractive summary of the article can significantly improve the generation of both types of keyphrases that are either present or absent from the text. Experimental results with three widely used models for keyphrase generation along with one of the latest transformer models suitable for longer documents, Longformer Encoder-Decoder (LED) validate the observation. We also present a new large-scale scholarly dataset FullTextKP for keyphrase generation. Unlike prior large-scale datasets, FullTextKP includes the full text of the articles along with the title and abstract. We release the source code at https://github.com/kgarg8/FullTextKP.
Recently, deep learning has attracted more and more attention in phase unwrapping of fringe projection three-dimensional (3D) measurement, with the aim to improve the performance leveraging the powerful Convolutional Neural Network (CNN) models. In this paper, for the first time (to the best of our knowledge), we introduce the Transformer into the phase unwrapping which is different from CNN and propose Hformer model dedicated to phase unwrapping via fringe order prediction. The proposed model has a hybrid CNN-Transformer architecture that is mainly composed of backbone, encoder and decoder to take advantage of both CNN and Transformer. Encoder and decoder with cross attention are designed for the fringe order prediction. Experimental results show that the proposed Hformer model achieves better performance in fringe order prediction compared with the CNN models such as U-Net and DCNN. Moreover, ablation study on Hformer is made to verify the improved feature pyramid networks (FPN) and testing strategy with flipping in the predicted fringe order. Our work opens an alternative way to deep learning based phase unwrapping methods, which are dominated by CNN in fringe projection 3D measurement.
The sparsely-gated Mixture of Experts (MoE) can magnify a network capacity with a little computational complexity. In this work, we investigate how multi-lingual Automatic Speech Recognition (ASR) networks can be scaled up with a simple routing algorithm in order to achieve better accuracy. More specifically, we apply the sparsely-gated MoE technique to two types of networks: Sequence-to-Sequence Transformer (S2S-T) and Transformer Transducer (T-T). We demonstrate through a set of ASR experiments on multiple language data that the MoE networks can reduce the relative word error rates by 16.3% and 4.6% with the S2S-T and T-T, respectively. Moreover, we thoroughly investigate the effect of the MoE on the T-T architecture in various conditions: streaming mode, non-streaming mode, the use of language ID and the label decoder with the MoE.
The increasing availability and accessibility of numerous overhead images allows us to estimate and assess the spatial arrangement of groups of geospatial target objects, which can benefit many applications, such as traffic monitoring and agricultural monitoring. Spatial arrangement estimation is the process of identifying the areas which contain the desired objects in overhead images. Traditional supervised object detection approaches can estimate accurate spatial arrangement but require large amounts of bounding box annotations. Recent semi-supervised clustering approaches can reduce manual labeling but still require annotations for all object categories in the image. This paper presents the target-guided generative model (TGGM), under the Variational Auto-encoder (VAE) framework, which uses Gaussian Mixture Models (GMM) to estimate the distributions of both hidden and decoder variables in VAE. Modeling both hidden and decoder variables by GMM reduces the required manual annotations significantly for spatial arrangement estimation. Unlike existing approaches that the training process can only update the GMM as a whole in the optimization iterations (e.g., a "minibatch"), TGGM allows the update of individual GMM components separately in the same optimization iteration. Optimizing GMM components separately allows TGGM to exploit the semantic relationships in spatial data and requires only a few labels to initiate and guide the generative process. Our experiments shows that TGGM achieves results comparable to the state-of-the-art semi-supervised methods and outperforms unsupervised methods by 10% based on the $F_{1}$ scores, while requiring significantly fewer labeled data.
Multi-label image classification is about predicting a set of class labels that can be considered as orderless sequential data. Transformers process the sequential data as a whole, therefore they are inherently good at set prediction. The first vision-based transformer model, which was proposed for the object detection task introduced the concept of object queries. Object queries are learnable positional encodings that are used by attention modules in decoder layers to decode the object classes or bounding boxes using the region of interests in an image. However, inputting the same set of object queries to different decoder layers hinders the training: it results in lower performance and delays convergence. In this paper, we propose the usage of primal object queries that are only provided at the start of the transformer decoder stack. In addition, we improve the mixup technique proposed for multi-label classification. The proposed transformer model with primal object queries improves the state-of-the-art class wise F1 metric by 2.1% and 1.8%; and speeds up the convergence by 79.0% and 38.6% on MS-COCO and NUS-WIDE datasets respectively.
Sarcasm is a pervading linguistic phenomenon and highly challenging to explain due to its subjectivity, lack of context and deeply-felt opinion. In the multimodal setup, sarcasm is conveyed through the incongruity between the text and visual entities. Although recent approaches deal with sarcasm as a classification problem, it is unclear why an online post is identified as sarcastic. Without proper explanation, end users may not be able to perceive the underlying sense of irony. In this paper, we propose a novel problem -- Multimodal Sarcasm Explanation (MuSE) -- given a multimodal sarcastic post containing an image and a caption, we aim to generate a natural language explanation to reveal the intended sarcasm. To this end, we develop MORE, a new dataset with explanation of 3510 sarcastic multimodal posts. Each explanation is a natural language (English) sentence describing the hidden irony. We benchmark MORE by employing a multimodal Transformer-based architecture. It incorporates a cross-modal attention in the Transformer's encoder which attends to the distinguishing features between the two modalities. Subsequently, a BART-based auto-regressive decoder is used as the generator. Empirical results demonstrate convincing results over various baselines (adopted for MuSE) across five evaluation metrics. We also conduct human evaluation on predictions and obtain Fleiss' Kappa score of 0.4 as a fair agreement among 25 evaluators.
Brain tumor segmentation is essential for the diagnosis and prognosis of patients with gliomas. The brain tumor segmentation challenge has continued to provide a great source of data to develop automatic algorithms to perform the task. This paper describes our contribution to the 2021 competition. We developed our methods based on nn-UNet, the winning entry of last year competition. We experimented with several modifications, including using a larger network, replacing batch normalization with group normalization, and utilizing axial attention in the decoder. Internal 5-fold cross validation as well as online evaluation from the organizers showed the effectiveness of our approach, with minor improvement in quantitative metrics when compared to the baseline. The proposed models won first place in the final ranking on unseen test data. The codes, pretrained weights, and docker image for the winning submission are publicly available at https://github.com/rixez/Brats21_KAIST_MRI_Lab
Although detection with Transformer (DETR) is increasingly popular, its global attention modeling requires an extremely long training period to optimize and achieve promising detection performance. Alternative to existing studies that mainly develop advanced feature or embedding designs to tackle the training issue, we point out that the Region-of-Interest (RoI) based detection refinement can easily help mitigate the difficulty of training for DETR methods. Based on this, we introduce a novel REcurrent Glimpse-based decOder (REGO) in this paper. In particular, the REGO employs a multi-stage recurrent processing structure to help the attention of DETR gradually focus on foreground objects more accurately. In each processing stage, visual features are extracted as glimpse features from RoIs with enlarged bounding box areas of detection results from the previous stage. Then, a glimpse-based decoder is introduced to provide refined detection results based on both the glimpse features and the attention modeling outputs of the previous stage. In practice, REGO can be easily embedded in representative DETR variants while maintaining their fully end-to-end training and inference pipelines. In particular, REGO helps Deformable DETR achieve 44.8 AP on the MSCOCO dataset with only 36 training epochs, compared with the first DETR and the Deformable DETR that require 500 and 50 epochs to achieve comparable performance, respectively. Experiments also show that REGO consistently boosts the performance of different DETR detectors by up to 7% relative gain at the same setting of 50 training epochs. Code is available via https://github.com/zhechen/Deformable-DETR-REGO.
We propose a method for learning the posture and structure of agents from unlabelled behavioral videos. Starting from the observation that behaving agents are generally the main sources of movement in behavioral videos, our method, Behavioral Keypoint Discovery (B-KinD), uses an encoder-decoder architecture with a geometric bottleneck to reconstruct the spatiotemporal difference between video frames. By focusing only on regions of movement, our approach works directly on input videos without requiring manual annotations. Experiments on a variety of agent types (mouse, fly, human, jellyfish, and trees) demonstrate the generality of our approach and reveal that our discovered keypoints represent semantically meaningful body parts, which achieve state-of-the-art performance on keypoint regression among self-supervised methods. Additionally, B-KinD achieve comparable performance to supervised keypoints on downstream tasks, such as behavior classification, suggesting that our method can dramatically reduce model training costs vis-a-vis supervised methods.
Vision transformer architectures have been demonstrated to work very effectively for image classification tasks. Efforts to solve more challenging vision tasks with transformers rely on convolutional backbones for feature extraction. In this paper we investigate the use of a pure transformer architecture (i.e., one with no CNN backbone) for the problem of 2D body pose estimation. We evaluate two ViT architectures on the COCO dataset. We demonstrate that using an encoder-decoder transformer architecture yields state of the art results on this estimation problem.
The aim of this work is to investigate the impact of crossmodal self-supervised pre-training for speech reconstruction (video-to-audio) by leveraging the natural co-occurrence of audio and visual streams in videos. We propose LipSound2 which consists of an encoder-decoder architecture and location-aware attention mechanism to map face image sequences to mel-scale spectrograms directly without requiring any human annotations. The proposed LipSound2 model is firstly pre-trained on $\sim$2400h multi-lingual (e.g. English and German) audio-visual data (VoxCeleb2). To verify the generalizability of the proposed method, we then fine-tune the pre-trained model on domain-specific datasets (GRID, TCD-TIMIT) for English speech reconstruction and achieve a significant improvement on speech quality and intelligibility compared to previous approaches in speaker-dependent and -independent settings. In addition to English, we conduct Chinese speech reconstruction on the CMLR dataset to verify the impact on transferability. Lastly, we train the cascaded lip reading (video-to-text) system by fine-tuning the generated audios on a pre-trained speech recognition system and achieve state-of-the-art performance on both English and Chinese benchmark datasets.
This paper presents a novel approach for the joint design of a reconfigurable intelligent surface (RIS) and a transmitter-receiver pair that are trained together as a set of deep neural networks (DNNs) to optimize the end-to-end communication performance at the receiver. The RIS is a software-defined array of unit cells that can be controlled in terms of the scattering and reflection profiles to focus the incoming signals from the transmitter to the receiver. The benefit of the RIS is to improve the coverage and spectral efficiency for wireless communications by overcoming physical obstructions of the line-of-sight (LoS) links. The selection process of the RIS beam codeword (out of a pre-defined codebook) is formulated as a DNN, while the operations of the transmitter-receiver pair are modeled as two DNNs, one for the encoder (at the transmitter) and the other one for the decoder (at the receiver) of an autoencoder, by accounting for channel effects including those induced by the RIS in between. The underlying DNNs are jointly trained to minimize the symbol error rate at the receiver. Numerical results show that the proposed design achieves major gains in error performance with respect to various baseline schemes, where no RIS is used or the selection of the RIS beam is separated from the design of the transmitter-receiver pair.
Unsupervised Zero-Shot Voice Conversion (VC) aims to modify the speaker characteristic of an utterance to match an unseen target speaker without relying on parallel training data. Recently, self-supervised learning of speech representation has been shown to produce useful linguistic units without using transcripts, which can be directly passed to a VC model. In this paper, we showed that high-quality audio samples can be achieved by using a length resampling decoder, which enables the VC model to work in conjunction with different linguistic feature extractors and vocoders without requiring them to operate on the same sequence length. We showed that our method can outperform many baselines on the VCTK dataset. Without modifying the architecture, we further demonstrated that a) using pairs of different audio segments from the same speaker, b) adding a cycle consistency loss, and c) adding a speaker classification loss can help to learn a better speaker embedding. Our model trained on LibriTTS using these techniques achieves the best performance, producing audio samples transferred well to the target speaker's voice, while preserving the linguistic content that is comparable with actual human utterances in terms of Character Error Rate.
Sign language translation as a kind of technology with profound social significance has attracted growing researchers' interest in recent years. However, the existing sign language translation methods need to read all the videos before starting the translation, which leads to a high inference latency and also limits their application in real-life scenarios. To solve this problem, we propose SimulSLT, the first end-to-end simultaneous sign language translation model, which can translate sign language videos into target text concurrently. SimulSLT is composed of a text decoder, a boundary predictor, and a masked encoder. We 1) use the wait-k strategy for simultaneous translation. 2) design a novel boundary predictor based on the integrate-and-fire module to output the gloss boundary, which is used to model the correspondence between the sign language video and the gloss. 3) propose an innovative re-encode method to help the model obtain more abundant contextual information, which allows the existing video features to interact fully. The experimental results conducted on the RWTH-PHOENIX-Weather 2014T dataset show that SimulSLT achieves BLEU scores that exceed the latest end-to-end non-simultaneous sign language translation model while maintaining low latency, which proves the effectiveness of our method.