Inferring the structure of 3D scenes from 2D observations is a fundamental challenge in computer vision. Recently popularized approaches based on neural scene representations have achieved tremendous impact and have been applied across a variety of applications. One of the major remaining challenges in this space is training a single model which can provide latent representations which effectively generalize beyond a single scene. Scene Representation Transformer (SRT) has shown promise in this direction, but scaling it to a larger set of diverse scenes is challenging and necessitates accurately posed ground truth data. To address this problem, we propose RUST (Really Unposed Scene representation Transformer), a pose-free approach to novel view synthesis trained on RGB images alone. Our main insight is that one can train a Pose Encoder that peeks at the target image and learns a latent pose embedding which is used by the decoder for view synthesis. We perform an empirical investigation into the learned latent pose structure and show that it allows meaningful test-time camera transformations and accurate explicit pose readouts. Perhaps surprisingly, RUST achieves similar quality as methods which have access to perfect camera pose, thereby unlocking the potential for large-scale training of amortized neural scene representations.
Holistic 3D scene understanding entails estimation of both layout configuration and object geometry in a 3D environment. Recent works have shown advances in 3D scene estimation from various input modalities (e.g., images, 3D scans), by leveraging 3D supervision (e.g., 3D bounding boxes or CAD models), for which collection at scale is expensive and often intractable. To address this shortcoming, we propose a new method to learn 3D scene priors of layout and shape without requiring any 3D ground truth. Instead, we rely on 2D supervision from multi-view RGB images. Our method represents a 3D scene as a latent vector, from which we can progressively decode to a sequence of objects characterized by their class categories, 3D bounding boxes, and meshes. With our trained autoregressive decoder representing the scene prior, our method facilitates many downstream applications, including scene synthesis, interpolation, and single-view reconstruction. Experiments on 3D-FRONT and ScanNet show that our method outperforms state of the art in single-view reconstruction, and achieves state-of-the-art results in scene synthesis against baselines which require for 3D supervision.
Pooling is essentially an operation from the field of Mathematical Morphology, with max pooling as a limited special case. The more general setting of MorphPooling greatly extends the tool set for building neural networks. In addition to pooling operations, encoder-decoder networks used for pixel-level predictions also require unpooling. It is common to combine unpooling with convolution or deconvolution for up-sampling. However, using its morphological properties, unpooling can be generalised and improved. Extensive experimentation on two tasks and three large-scale datasets shows that morphological pooling and unpooling lead to improved predictive performance at much reduced parameter counts.
This paper explores the multi-scale aggregation strategy for scene text detection in natural images. We present the Aggregated Text TRansformer(ATTR), which is designed to represent texts in scene images with a multi-scale self-attention mechanism. Starting from the image pyramid with multiple resolutions, the features are first extracted at different scales with shared weight and then fed into an encoder-decoder architecture of Transformer. The multi-scale image representations are robust and contain rich information on text contents of various sizes. The text Transformer aggregates these features to learn the interaction across different scales and improve text representation. The proposed method detects scene texts by representing each text instance as an individual binary mask, which is tolerant of curve texts and regions with dense instances. Extensive experiments on public scene text detection datasets demonstrate the effectiveness of the proposed framework.
We present a large-scale facial UV-texture dataset that contains over 50,000 high-quality texture UV-maps with even illuminations, neutral expressions, and cleaned facial regions, which are desired characteristics for rendering realistic 3D face models under different lighting conditions. The dataset is derived from a large-scale face image dataset namely FFHQ, with the help of our fully automatic and robust UV-texture production pipeline. Our pipeline utilizes the recent advances in StyleGAN-based facial image editing approaches to generate multi-view normalized face images from single-image inputs. An elaborated UV-texture extraction, correction, and completion procedure is then applied to produce high-quality UV-maps from the normalized face images. Compared with existing UV-texture datasets, our dataset has more diverse and higher-quality texture maps. We further train a GAN-based texture decoder as the nonlinear texture basis for parametric fitting based 3D face reconstruction. Experiments show that our method improves the reconstruction accuracy over state-of-the-art approaches, and more importantly, produces high-quality texture maps that are ready for realistic renderings. The dataset, code, and pre-trained texture decoder are publicly available at https://github.com/csbhr/FFHQ-UV.
Automatic detection of anomalies such as weapons or threat objects in baggage security, or detecting impaired items in industrial production is an important computer vision task demanding high efficiency and accuracy. Most of the available data in the anomaly detection task is imbalanced as the number of positive/anomalous instances is sparse. Inadequate availability of the data makes training of a deep neural network architecture for anomaly detection challenging. This paper proposes a novel Generative Adversarial Network (GAN) based model for anomaly detection. It uses normal (non-anomalous) images to learn about the normality based on which it detects if an input image contains an anomalous/threat object. The proposed model uses a generator with an encoder-decoder network having dense convolutional skip connections for enhanced reconstruction and to capture the data distribution. A self-attention augmented discriminator is used having the ability to check the consistency of detailed features even in distant portions. We use spectral normalisation to facilitate stable and improved training of the GAN. Experiments are performed on three datasets, viz. CIFAR-10, MVTec AD (for industrial applications) and SIXray (for X-ray baggage security). On the MVTec AD and SIXray datasets, our model achieves an improvement of upto 21% and 4.6%, respectively
Text error correction aims to correct the errors in text sequences such as those typed by humans or generated by speech recognition models. Previous error correction methods usually take the source (incorrect) sentence as encoder input and generate the target (correct) sentence through the decoder. Since the error rate of the incorrect sentence is usually low (e.g., 10\%), the correction model can only learn to correct on limited error tokens but trivially copy on most tokens (correct tokens), which harms the effective training of error correction. In this paper, we argue that the correct tokens should be better utilized to facilitate effective training and then propose a simple yet effective masking strategy to achieve this goal. Specifically, we randomly mask out a part of the correct tokens in the source sentence and let the model learn to not only correct the original error tokens but also predict the masked tokens based on their context information. Our method enjoys several advantages: 1) it alleviates trivial copy; 2) it leverages effective training signals from correct tokens; 3) it is a plug-and-play module and can be applied to different models and tasks. Experiments on spelling error correction and speech recognition error correction on Mandarin datasets and grammar error correction on English datasets with both autoregressive and non-autoregressive generation models show that our method improves the correction accuracy consistently.
GPS trajectories are the essential foundations for many trajectory-based applications, such as travel time estimation, traffic prediction and trajectory similarity measurement. Most applications require a large amount of high sample rate trajectories to achieve a good performance. However, many real-life trajectories are collected with low sample rate due to energy concern or other constraints.We study the task of trajectory recovery in this paper as a means for increasing the sample rate of low sample trajectories. Currently, most existing works on trajectory recovery follow a sequence-to-sequence diagram, with an encoder to encode a trajectory and a decoder to recover real GPS points in the trajectory. However, these works ignore the topology of road network and only use grid information or raw GPS points as input. Therefore, the encoder model is not able to capture rich spatial information of the GPS points along the trajectory, making the prediction less accurate and lack spatial consistency. In this paper, we propose a road network enhanced transformer-based framework, namely RNTrajRec, for trajectory recovery. RNTrajRec first uses a graph model, namely GridGNN, to learn the embedding features of each road segment. It next develops a spatial-temporal transformer model, namely GPSFormer, to learn rich spatial and temporal features along with a Sub-Graph Generation module to capture the spatial features for each GPS point in the trajectory. It finally forwards the outputs of encoder model into a multi-task decoder model to recover the missing GPS points. Extensive experiments based on three large-scale real-life trajectory datasets confirm the effectiveness of our approach.
Nowadays, semantic segmentation methods for systems in road scene have a great demand. Most existing methods focus on high accuracy with low inference speed. And some approaches emphasize on speed, significantly sacrificing model accuracy. To make a trade-off between accuracy and inference speed, we propose a real-time network for semantic segmentation titled Factorized and Regular Network (FRNet), which employs an asymmetric encoder-decoder architecture with Factorized and Regular (FR) blocks. Our method achieves 70.4% mIoU on the Cityscapes test set with 1 million parameters at a speed of 127 frames per second (FPS) on a single Titan Xp at a resolution of 512×1024 . We evaluate FRNet on Cityscapes, Camvid, Kitti, and Gatech datasets to identify that our network stands out from other state-of-the-art networks.
In recent works on semantic segmentation, there has been a significant focus on designing and integrating transformer-based encoders. However, less attention has been given to transformer-based decoders. We emphasize that the decoder stage is equally vital as the encoder in achieving superior segmentation performance. It disentangles and refines high-level cues, enabling precise object boundary delineation at the pixel level. In this paper, we introduce a novel transformer-based decoder called MUSTER, which seamlessly integrates with hierarchical encoders and consistently delivers high-quality segmentation results, regardless of the encoder architecture. Furthermore, we present a variant of MUSTER that reduces FLOPS while maintaining performance. MUSTER incorporates carefully designed multi-head skip attention (MSKA) units and introduces innovative upsampling operations. The MSKA units enable the fusion of multi-scale features from the encoder and decoder, facilitating comprehensive information integration. The upsampling operation leverages encoder features to enhance object localization and surpasses traditional upsampling methods, improving mIoU (mean Intersection over Union) by 0.4% to 3.2%. On the challenging ADE20K dataset, our best model achieves a single-scale mIoU of 50.23 and a multi-scale mIoU of 51.88, which is on-par with the current state-of-the-art model. Remarkably, we achieve this while significantly reducing the number of FLOPs by 61.3%. Our source code and models are publicly available at: https://github.com/shiwt03/MUSTER.
Chest X-ray is one of the most popular medical imaging modalities due to its accessibility and effectiveness. However, there is a chronic shortage of well-trained radiologists who can interpret these images and diagnose the patient's condition. Therefore, automated radiology report generation can be a very helpful tool in clinical practice. A typical report generation workflow consists of two main steps: (i) encoding the image into a latent space and (ii) generating the text of the report based on the latent image embedding. Many existing report generation techniques use a standard convolutional neural network (CNN) architecture for image encoding followed by a Transformer-based decoder for medical text generation. In most cases, CNN and the decoder are trained jointly in an end-to-end fashion. In this work, we primarily focus on understanding the relative importance of encoder and decoder components. Towards this end, we analyze four different image encoding approaches: direct, fine-grained, CLIP-based, and Cluster-CLIP-based encodings in conjunction with three different decoders on the large-scale MIMIC-CXR dataset. Among these encoders, the cluster CLIP visual encoder is a novel approach that aims to generate more discriminative and explainable representations. CLIP-based encoders produce comparable results to traditional CNN-based encoders in terms of NLP metrics, while fine-grained encoding outperforms all other encoders both in terms of NLP and clinical accuracy metrics, thereby validating the importance of image encoder to effectively extract semantic information. GitHub repository: https://github.com/mudabek/encoding-cxr-report-gen
Large sequence to sequence models for tasks such as Neural Machine Translation (NMT) are usually trained over hundreds of millions of samples. However, training is just the origin of a model's life-cycle. Real-world deployments of models require further behavioral adaptations as new requirements emerge or shortcomings become known. Typically, in the space of model behaviors, behavior deletion requests are addressed through model retrainings whereas model finetuning is done to address behavior addition requests, both procedures being instances of data-based model intervention. In this work, we present a preliminary study investigating rank-one editing as a direct intervention method for behavior deletion requests in encoder-decoder transformer models. We propose four editing tasks for NMT and show that the proposed editing algorithm achieves high efficacy, while requiring only a single instance of positive example to fix an erroneous (negative) model behavior.
In this paper, we provide the observation that too few queries assigned as positive samples in DETR with one-to-one set matching leads to sparse supervision on the encoder's output which considerably hurt the discriminative feature learning of the encoder and vice visa for attention learning in the decoder. To alleviate this, we present a novel collaborative hybrid assignments training scheme, namely $\mathcal{C}$o-DETR, to learn more efficient and effective DETR-based detectors from versatile label assignment manners. This new training scheme can easily enhance the encoder's learning ability in end-to-end detectors by training the multiple parallel auxiliary heads supervised by one-to-many label assignments such as ATSS and Faster RCNN. In addition, we conduct extra customized positive queries by extracting the positive coordinates from these auxiliary heads to improve the training efficiency of positive samples in the decoder. In inference, these auxiliary heads are discarded and thus our method introduces no additional parameters and computational cost to the original detector while requiring no hand-crafted non-maximum suppression (NMS). We conduct extensive experiments to evaluate the effectiveness of the proposed approach on DETR variants, including DAB-DETR, Deformable-DETR, and DINO-Deformable-DETR. The state-of-the-art DINO-Deformable-DETR with Swin-L can be improved from 58.5% to 59.5% AP on COCO val. Surprisingly, incorporated with ViT-L backbone, we achieve 66.0% AP on COCO test-dev and 67.9% AP on LVIS val, outperforming previous methods by clear margins with much fewer model sizes. Codes are available at \url{https://github.com/Sense-X/Co-DETR}.
In many vision and robotics applications, it is common that the captured objects are represented by very few points. Most of the existing completion methods are designed for partial point clouds with many points, and they perform poorly or even fail completely in the case of few points. However, due to the lack of detail information, completing objects from few points faces a huge challenge. Inspired by the successful applications of GAN and Transformers in the image-based vision task, we introduce GAN and Transformer techniques to address the above problem. Firstly, the end-to-end encoder-decoder network with Transformers and the Wasserstein GAN with Transformer are pre-trained, and then the overall network is fine-tuned. Experimental results on the ShapeNet dataset show that our method can not only improve the completion performance for many input points, but also keep stable for few input points. Our source code is available at https://github.com/WxfQjh/Stability-point-recovery.git.
One of the challenges of developing a summarization model arises from the difficulty in measuring the factual inconsistency of the generated text. In this study, we reinterpret the decoder overconfidence-regularizing objective suggested in (Miao et al., 2021) as a hallucination risk measurement to better estimate the quality of generated summaries. We propose a reference-free metric, HaRiM+, which only requires an off-the-shelf summarization model to compute the hallucination risk based on token likelihoods. Deploying it requires no additional training of models or ad-hoc modules, which usually need alignment to human judgments. For summary-quality estimation, HaRiM+ records state-of-the-art correlation to human judgment on three summary-quality annotation sets: FRANK, QAGS, and SummEval. We hope that our work, which merits the use of summarization models, facilitates the progress of both automated evaluation and generation of summary.
We present AERO, a audio super-resolution model that processes speech and music signals in the spectral domain. AERO is based on an encoder-decoder architecture with U-Net like skip connections. We optimize the model using both time and frequency domain loss functions. Specifically, we consider a set of reconstruction losses together with perceptual ones in the form of adversarial and feature discriminator loss functions. To better handle phase information the proposed method operates over the complex-valued spectrogram using two separate channels. Unlike prior work which mainly considers low and high frequency concatenation for audio super-resolution, the proposed method directly predicts the full frequency range. We demonstrate high performance across a wide range of sample rates considering both speech and music. AERO outperforms the evaluated baselines considering Log-Spectral Distance, ViSQOL, and the subjective MUSHRA test. Audio samples and code are available at https://pages.cs.huji.ac.il/adiyoss-lab/aero
Using a text description as prompt to guide the generation of text or images (e.g., GPT-3 or DALLE-2) has drawn wide attention recently. Beyond text and image generation, in this work, we explore the possibility of utilizing text descriptions to guide speech synthesis. Thus, we develop a text-to-speech (TTS) system (dubbed as PromptTTS) that takes a prompt with both style and content descriptions as input to synthesize the corresponding speech. Specifically, PromptTTS consists of a style encoder and a content encoder to extract the corresponding representations from the prompt, and a speech decoder to synthesize speech according to the extracted style and content representations. Compared with previous works in controllable TTS that require users to have acoustic knowledge to understand style factors such as prosody and pitch, PromptTTS is more user-friendly since text descriptions are a more natural way to express speech style (e.g., ''A lady whispers to her friend slowly''). Given that there is no TTS dataset with prompts, to benchmark the task of PromptTTS, we construct and release a dataset containing prompts with style and content information and the corresponding speech. Experiments show that PromptTTS can generate speech with precise style control and high speech quality. Audio samples and our dataset are publicly available.
The task of event extraction (EE) aims to find the events and event-related argument information from the text and represent them in a structured format. Most previous works try to solve the problem by separately identifying multiple substructures and aggregating them to get the complete event structure. The problem with the methods is that it fails to identify all the interdependencies among the event participants (event-triggers, arguments, and roles). In this paper, we represent each event record in a unique tuple format that contains trigger phrase, trigger type, argument phrase, and corresponding role information. Our proposed pointer network-based encoder-decoder model generates an event tuple in each time step by exploiting the interactions among event participants and presenting a truly end-to-end solution to the EE task. We evaluate our model on the ACE2005 dataset, and experimental results demonstrate the effectiveness of our model by achieving competitive performance compared to the state-of-the-art methods.
Multiphysics problems such as multicomponent diffusion, phase transformations in multiphase systems and alloy solidification involve numerical solution of a coupled system of nonlinear partial differential equations (PDEs). Numerical solutions of these PDEs using mesh-based methods require spatiotemporal discretization of these equations. Hence, the numerical solutions are often sensitive to discretization parameters and may have inaccuracies (resulting from grid-based approximations). Moreover, choice of finer mesh for higher accuracy make these methods computationally expensive. Neural network-based PDE solvers are emerging as robust alternatives to conventional numerical methods because these use machine learnable structures that are grid-independent, fast and accurate. However, neural network based solvers require large amount of training data, thus affecting their generalizabilty and scalability. These concerns become more acute for coupled systems of time-dependent PDEs. To address these issues, we develop a new neural network based framework that uses encoder-decoder based conditional Generative Adversarial Networks with ConvLSTM layers to solve a system of Cahn-Hilliard equations. These equations govern microstructural evolution of a ternary alloy undergoing spinodal decomposition when quenched inside a three-phase miscibility gap. We show that the trained models are mesh and scale-independent, thereby warranting application as effective neural operators.
Weakly Supervised Semantic Segmentation (WSSS) is a challenging task aiming to learn the segmentation labels from class-level labels. In the literature, exploiting the information obtained from Class Activation Maps (CAMs) is widely used for WSSS studies. However, as CAMs are obtained from a classification network, they are interested in the most discriminative parts of the objects, producing non-complete prior information for segmentation tasks. In this study, to obtain more coherent CAMs with segmentation labels, we propose a framework that employs an iterative approach in a modified encoder-decoder-based segmentation model, which simultaneously supports classification and segmentation tasks. As no ground-truth segmentation labels are given, the same model also generates the pseudo-segmentation labels with the help of dense Conditional Random Fields (dCRF). As a result, the proposed framework becomes an iterative self-improved model. The experiments performed with DeepLabv3 and UNet models show a significant gain on the Pascal VOC12 dataset, and the DeepLabv3 application increases the current state-of-the-art metric by %2.5. The implementation associated with the experiments can be found: https://github.com/cenkbircanoglu/isim.