The autoencoder model uses an encoder to map data samples to a lower dimensional latent space and then a decoder to map the latent space representations back to the data space. Implicitly, it relies on the encoder to approximate the inverse of the decoder network, so that samples can be mapped to and back from the latent space faithfully. This approximation may lead to sub-optimal latent space representations. In this work, we investigate a decoder-only method that uses gradient flow to encode data samples in the latent space. The gradient flow is defined based on a given decoder and aims to find the optimal latent space representation for any given sample through optimisation, eliminating the need of an approximate inversion through an encoder. Implementing gradient flow through ordinary differential equations (ODE), we leverage the adjoint method to train a given decoder. We further show empirically that the costly integrals in the adjoint method may not be entirely necessary. Additionally, we propose a $2^{nd}$ order ODE variant to the method, which approximates Nesterov's accelerated gradient descent, with faster convergence per iteration. Commonly used ODE solvers can be quite sensitive to the integration step-size depending on the stiffness of the ODE. To overcome the sensitivity for gradient flow encoding, we use an adaptive solver that prioritises minimising loss at each integration step. We assess the proposed method in comparison to the autoencoding model. In our experiments, GFE showed a much higher data-efficiency than the autoencoding model, which can be crucial for data scarce applications.
The paper introduces a novel, holistic approach for robust Screen-Camera Communication (SCC), where video content on a screen is visually encoded in a human-imperceptible fashion and decoded by a camera capturing images of such screen content. We first show that state-of-the-art SCC techniques have two key limitations for in-the-wild deployment: (a) the decoding accuracy drops rapidly under even modest screen extraction errors from the captured images, and (b) they generate perceptible flickers on common refresh rate screens even with minimal modulation of pixel intensity. To overcome these challenges, we introduce DeepLight, a system that incorporates machine learning (ML) models in the decoding pipeline to achieve humanly-imperceptible, moderately high SCC rates under diverse real-world conditions. Deep-Light's key innovation is the design of a Deep Neural Network (DNN) based decoder that collectively decodes all the bits spatially encoded in a display frame, without attempting to precisely isolate the pixels associated with each encoded bit. In addition, DeepLight supports imperceptible encoding by selectively modulating the intensity of only the Blue channel, and provides reasonably accurate screen extraction (IoU values >= 83%) by using state-of-the-art object detection DNN pipelines. We show that a fully functional DeepLight system is able to robustly achieve high decoding accuracy (frame error rate < 0.2) and moderately-high data goodput (>=0.95Kbps) using a human-held smartphone camera, even over larger screen-camera distances (approx =2m).
The spatial attention is a straightforward approach to enhance the performance for remote sensing image captioning. However, conventional spatial attention approaches consider only the attention distribution on one fixed coarse grid, resulting in the semantics of tiny objects can be easily ignored or disturbed during the visual feature extraction. Worse still, the fixed semantic level of conventional spatial attention limits the image understanding in different levels and perspectives, which is critical for tackling the huge diversity in remote sensing images. To address these issues, we propose a remote sensing image caption generator with instance-awareness and cross-hierarchy attention. 1) The instances awareness is achieved by introducing a multi-level feature architecture that contains the visual information of multi-level instance-possible regions and their surroundings. 2) Moreover, based on this multi-level feature extraction, a cross-hierarchy attention mechanism is proposed to prompt the decoder to dynamically focus on different semantic hierarchies and instances at each time step. The experimental results on public datasets demonstrate the superiority of proposed approach over existing methods.
This paper describes the submission to the IWSLT 2021 offline speech translation task by the UPC Machine Translation group. The task consists of building a system capable of translating English audio recordings extracted from TED talks into German text. Submitted systems can be either cascade or end-to-end and use a custom or given segmentation. Our submission is an end-to-end speech translation system, which combines pre-trained models (Wav2Vec 2.0 and mBART) with coupling modules between the encoder and decoder, and uses an efficient fine-tuning technique, which trains only 20% of its total parameters. We show that adding an Adapter to the system and pre-training it, can increase the convergence speed and the final result, with which we achieve a BLEU score of 27.3 on the MuST-C test set. Our final model is an ensemble that obtains 28.22 BLEU score on the same set. Our submission also uses a custom segmentation algorithm that employs pre-trained Wav2Vec 2.0 for identifying periods of untranscribable text and can bring improvements of 2.5 to 3 BLEU score on the IWSLT 2019 test set, as compared to the result with the given segmentation.
Powerful deep learning tools, such as convolutional neural networks (CNN), are able to learn the input-output relationships of large complicated systems directly from data. Encoder-decoder deep CNNs are able to extract features directly from images, mix them with scalar inputs within a general low-dimensional latent space, and then generate new complex 2D outputs which represent complex physical phenomenon. One important challenge faced by deep learning methods is large non-stationary systems whose characteristics change quickly with time for which re-training is not feasible. In this paper we present a method for adaptive tuning of the low-dimensional latent space of deep encoder-decoder style CNNs based on real-time feedback to quickly compensate for unknown and fast distribution shifts. We demonstrate our approach for predicting the properties of a time-varying charged particle beam in a particle accelerator whose components (accelerating electric fields and focusing magnetic fields) are also quickly changing with time.
Decoder diversity is a powerful error correction framework in which a collection of decoders collaboratively correct a set of error patterns otherwise uncorrectable by any individual decoder. In this paper, we propose a new approach to design the decoder diversity of finite alphabet iterative decoders (FAIDs) for Low-Density Parity Check (LDPC) codes over the binary symmetric channel (BSC), for the purpose of lowering the error floor while guaranteeing the waterfall performance. The proposed decoder diversity is achieved by training a recurrent quantized neural network (RQNN) to learn/design FAIDs. We demonstrated for the first time that a machine-learned decoder can surpass in performance a man-made decoder of the same complexity. As RQNNs can model a broad class of FAIDs, they are capable of learning an arbitrary FAID. To provide sufficient knowledge of the error floor to the RQNN, the training sets are constructed by sampling from the set of most problematic error patterns - trapping sets. In contrast to the existing methods that use the cross-entropy function as the loss function, we introduce a frame-error-rate (FER) based loss function to train the RQNN with the objective of correcting specific error patterns rather than reducing the bit error rate (BER). The examples and simulation results show that the RQNN-aided decoder diversity increases the error correction capability of LDPC codes and lowers the error floor.
Transformers and variational autoencoders (VAE) have been extensively employed for symbolic (e.g., MIDI) domain music generation. While the former boast an impressive capability in modeling long sequences, the latter allow users to willingly exert control over different parts (e.g., bars) of the music to be generated. In this paper, we are interested in bringing the two together to construct a single model that exhibits both strengths. The task is split into two steps. First, we equip Transformer decoders with the ability to accept segment-level, time-varying conditions during sequence generation. Subsequently, we combine the developed and tested in-attention decoder with a Transformer encoder, and train the resulting MuseMorphose model with the VAE objective to achieve style transfer of long pop piano pieces, in which users can specify musical attributes including rhythmic intensity and polyphony (i.e., harmonic fullness) they desire, down to the bar level. Experiments show that MuseMorphose outperforms recurrent neural network (RNN) based baselines on numerous widely-used metrics for style transfer tasks.
Template-based discriminative trackers are currently the dominant tracking methods due to their robustness and accuracy, and the Siamese-network-based methods that depend on cross-correlation operation between features extracted from template and search images show the state-of-the-art tracking performance. However, general cross-correlation operation can only obtain relationship between local patches in two feature maps. In this paper, we propose a novel tracker network based on a powerful attention mechanism called Transformer encoder-decoder architecture to gain global and rich contextual interdependencies. In this new architecture, features of the template image is processed by a self-attention module in the encoder part to learn strong context information, which is then sent to the decoder part to compute cross-attention with the search image features processed by another self-attention module. In addition, we design the classification and regression heads using the output of Transformer to localize target based on shape-agnostic anchor. We extensively evaluate our tracker TrTr, on VOT2018, VOT2019, OTB-100, UAV, NfS, TrackingNet, and LaSOT benchmarks and our method performs favorably against state-of-the-art algorithms. Training code and pretrained models are available at https://github.com/tongtybj/TrTr.
Scene text recognition is a challenging task due to diverse variations of text instances in natural scene images. Conventional methods based on CNN-RNN-CTC or encoder-decoder with attention mechanism may not fully investigate stable and efficient feature representations for multi-oriented scene texts. In this paper, we propose a primitive representation learning method that aims to exploit intrinsic representations of scene text images. We model elements in feature maps as the nodes of an undirected graph. A pooling aggregator and a weighted aggregator are proposed to learn primitive representations, which are transformed into high-level visual text representations by graph convolutional networks. A Primitive REpresentation learning Network (PREN) is constructed to use the visual text representations for parallel decoding. Furthermore, by integrating visual text representations into an encoder-decoder model with the 2D attention mechanism, we propose a framework called PREN2D to alleviate the misalignment problem in attention-based methods. Experimental results on both English and Chinese scene text recognition tasks demonstrate that PREN keeps a balance between accuracy and efficiency, while PREN2D achieves state-of-the-art performance.
Multi-level features are important for saliency detection. Better combination and use of multi-level features with time information can greatly improve the accuracy of the video saliency model. In order to fully combine multi-level features and make it serve the video saliency model, we propose a 3D fully convolutional encoder-decoder architecture for video saliency detection, which combines scale, space and time information for video saliency modeling. The encoder extracts multi-scale temporal-spatial features from the input continuous video frames, and then constructs temporal-spatial feature pyramid through temporal-spatial convolution and top-down feature integration. The decoder performs hierarchical decoding of temporal-spatial features from different scales, and finally produces a saliency map from the integration of multiple video frames. Our model is simple yet effective, and can run in real time. We perform abundant experiments, and the results indicate that the well-designed structure can improve the precision of video saliency detection significantly. Experimental results on three purely visual video saliency benchmarks and six audio-video saliency benchmarks demonstrate that our method outperforms the existing state-of-the-art methods.
Currently, developments of deep learning techniques are providing instrumental to identify, classify, and quantify patterns in medical images. Segmentation is one of the important applications in medical image analysis. In this regard, U-Net is the predominant approach to medical image segmentation tasks. However, we found that those U-Net based models have limitations in several aspects, for example, millions of parameters in the U-Net consuming considerable computation resource and memory, lack of global information, and missing some tough objects. Therefore, we applied two modifications to improve the U-Net model: 1) designed and added the dilated channel-wise CNN module, 2) simplified the U shape network. Based on these two modifications, we proposed a novel light-weight architecture -- Channel-wise Feature Pyramid Network for Medicine (CFPNet-M). To evaluate our method, we selected five datasets with different modalities: thermography, electron microscopy, endoscopy, dermoscopy, and digital retinal images. And we compared its performance with several models having different parameter scales. This paper also involves our previous studies of DC-UNet and some commonly used light-weight neural networks. We applied the Tanimoto similarity instead of the Jaccard index for gray-level image measurements. By comparison, CFPNet-M achieves comparable segmentation results on all five medical datasets with only 0.65 million parameters, which is about 2% of U-Net, and 8.8 MB memory. Meanwhile, the inference speed can reach 80 FPS on a single RTX 2070Ti GPU with the 256 by 192 pixels input size.
Convolutional neural networks lack shift equivariance due to the presence of downsampling layers. In image classification, adaptive polyphase downsampling (APS-D) was recently proposed to make CNNs perfectly shift invariant. However, in networks used for image reconstruction tasks, it can not by itself restore shift equivariance. We address this problem by proposing adaptive polyphase upsampling (APS-U), a non-linear extension of conventional upsampling, which allows CNNs with symmetric encoder-decoder architecture (for example U-Net) to exhibit perfect shift equivariance. With MRI and CT reconstruction experiments, we show that networks containing APS-D/U layers exhibit state of the art equivariance performance without sacrificing on image reconstruction quality. In addition, unlike prior methods like data augmentation and anti-aliasing, the gains in equivariance obtained from APS-D/U also extend to images outside the training distribution.
In this paper, we propose a transformer based approach for visual grounding. Unlike previous proposal-and-rank frameworks that rely heavily on pretrained object detectors or proposal-free frameworks that upgrade an off-the-shelf one-stage detector by fusing textual embeddings, our approach is built on top of a transformer encoder-decoder and is independent of any pretrained detectors or word embedding models. Termed VGTR -- Visual Grounding with TRansformers, our approach is designed to learn semantic-discriminative visual features under the guidance of the textual description without harming their location ability. This information flow enables our VGTR to have a strong capability in capturing context-level semantics of both vision and language modalities, rendering us to aggregate accurate visual clues implied by the description to locate the interested object instance. Experiments show that our method outperforms state-of-the-art proposal-free approaches by a considerable margin on five benchmarks while maintaining fast inference speed.
Artifact removal and filtering methods are inevitable parts of video coding. On one hand, new codecs and compression standards come with advanced in-loop filters and on the other hand, displays are equipped with high capacity processing units for post-treatment of decoded videos. This paper proposes a Convolutional Neural Network (CNN)-based post-processing algorithm for intra and inter frames of Versatile Video Coding (VVC) coded streams. Depending on the frame type, this method benefits from normative prediction signal by feeding it as an additional input along with reconstructed signal and a Quantization Parameter (QP)-map to the CNN. Moreover, an optional Model Selection (MS) strategy is adopted to pick the best trained model among available ones at the encoder side and signal it to the decoder side. This MS strategy is applicable at both frame level and block level. The experiments under the Random Access (RA) configuration of the VVC Test Model (VTM-10.0) show that the proposed prediction-aware algorithm can bring an additional BD-BR gain of -1.3% compared to the method without the prediction information. Furthermore, the proposed MS scheme brings -0.5% more BD-BR gain on top of the prediction-aware method.
Deep neural networks often suffer from overconfidence which can be partly remedied by improved out-of-distribution detection. For this purpose, we propose a novel approach that allows for the generation of out-of-distribution datasets based on a given in-distribution dataset. This new dataset can then be used to improve out-of-distribution detection for the given dataset and machine learning task at hand. The samples in this dataset are with respect to the feature space close to the in-distribution dataset and therefore realistic and plausible. Hence, this dataset can also be used to safeguard neural networks, i.e., to validate the generalization performance. Our approach first generates suitable representations of an in-distribution dataset using an autoencoder and then transforms them using our novel proposed Soft Brownian Offset method. After transformation, the decoder part of the autoencoder allows for the generation of these implicit out-of-distribution samples. This newly generated dataset then allows for mixing with other datasets and thus improved training of an out-of-distribution classifier, increasing its performance. Experimentally, we show that our approach is promising for time series using synthetic data. Using our new method, we also show in a quantitative case study that we can improve the out-of-distribution detection for the MNIST dataset. Finally, we provide another case study on the synthetic generation of out-of-distribution trajectories, which can be used to validate trajectory prediction algorithms for automated driving.
Encoder-decoder models have made great progress on handwritten mathematical expression recognition recently. However, it is still a challenge for existing methods to assign attention to image features accurately. Moreover, those encoder-decoder models usually adopt RNN-based models in their decoder part, which makes them inefficient in processing long $\LaTeX{}$ sequences. In this paper, a transformer-based decoder is employed to replace RNN-based ones, which makes the whole model architecture very concise. Furthermore, a novel training strategy is introduced to fully exploit the potential of the transformer in bidirectional language modeling. Compared to several methods that do not use data augmentation, experiments demonstrate that our model improves the ExpRate of current state-of-the-art methods on CROHME 2014 by 2.23%. Similarly, on CROHME 2016 and CROHME 2019, we improve the ExpRate by 1.92% and 2.28% respectively.
Image captioning is one of the most challenging tasks in AI, which aims to automatically generate textual sentences for an image. Recent methods for image captioning follow encoder-decoder framework that transforms the sequence of salient regions in an image into natural language descriptions. However, these models usually lack the comprehensive understanding of the contextual interactions reflected on various visual relationships between objects. In this paper, we explore explicit and implicit visual relationships to enrich region-level representations for image captioning. Explicitly, we build semantic graph over object pairs and exploit gated graph convolutional networks (Gated GCN) to selectively aggregate local neighbors' information. Implicitly, we draw global interactions among the detected objects through region-based bidirectional encoder representations from transformers (Region BERT) without extra relational annotations. To evaluate the effectiveness and superiority of our proposed method, we conduct extensive experiments on Microsoft COCO benchmark and achieve remarkable improvements compared with strong baselines.
Despite over two decades of progress, imbalanced data is still considered a significant challenge for contemporary machine learning models. Modern advances in deep learning have magnified the importance of the imbalanced data problem. The two main approaches to address this issue are based on loss function modifications and instance resampling. Instance sampling is typically based on Generative Adversarial Networks (GANs), which may suffer from mode collapse. Therefore, there is a need for an oversampling method that is specifically tailored to deep learning models, can work on raw images while preserving their properties, and is capable of generating high quality, artificial images that can enhance minority classes and balance the training set. We propose DeepSMOTE - a novel oversampling algorithm for deep learning models. It is simple, yet effective in its design. It consists of three major components: (i) an encoder/decoder framework; (ii) SMOTE-based oversampling; and (iii) a dedicated loss function that is enhanced with a penalty term. An important advantage of DeepSMOTE over GAN-based oversampling is that DeepSMOTE does not require a discriminator, and it generates high-quality artificial images that are both information-rich and suitable for visual inspection. DeepSMOTE code is publicly available at: https://github.com/dd1github/DeepSMOTE
In this paper, we propose a two-stage deep learning framework called VoxelContext-Net for both static and dynamic point cloud compression. Taking advantages of both octree based methods and voxel based schemes, our approach employs the voxel context to compress the octree structured data. Specifically, we first extract the local voxel representation that encodes the spatial neighbouring context information for each node in the constructed octree. Then, in the entropy coding stage, we propose a voxel context based deep entropy model to compress the symbols of non-leaf nodes in a lossless way. Furthermore, for dynamic point cloud compression, we additionally introduce the local voxel representations from the temporal neighbouring point clouds to exploit temporal dependency. More importantly, to alleviate the distortion from the octree construction procedure, we propose a voxel context based 3D coordinate refinement method to produce more accurate reconstructed point cloud at the decoder side, which is applicable to both static and dynamic point cloud compression. The comprehensive experiments on both static and dynamic point cloud benchmark datasets(e.g., ScanNet and Semantic KITTI) clearly demonstrate the effectiveness of our newly proposed method VoxelContext-Net for 3D point cloud geometry compression.
Measuring lesion size is an important step to assess tumor growth and monitor disease progression and therapy response in oncology image analysis. Although it is tedious and highly time-consuming, radiologists have to work on this task by using RECIST criteria (Response Evaluation Criteria In Solid Tumors) routinely and manually. Even though lesion segmentation may be the more accurate and clinically more valuable means, physicians can not manually segment lesions as now since much more heavy laboring will be required. In this paper, we present a prior-guided dual-path network (PDNet) to segment common types of lesions throughout the whole body and predict their RECIST diameters accurately and automatically. Similar to [1], a click guidance from radiologists is the only requirement. There are two key characteristics in PDNet: 1) Learning lesion-specific attention matrices in parallel from the click prior information by the proposed prior encoder, named click-driven attention; 2) Aggregating the extracted multi-scale features comprehensively by introducing top-down and bottom-up connections in the proposed decoder, named dual-path connection. Experiments show the superiority of our proposed PDNet in lesion segmentation and RECIST diameter prediction using the DeepLesion dataset and an external test set. PDNet learns comprehensive and representative deep image features for our tasks and produces more accurate results on both lesion segmentation and RECIST diameter prediction.