For video captioning, "pre-training and fine-tuning" has become a de facto paradigm, where ImageNet Pre-training (INP) is usually used to help encode the video content, and a task-oriented network is fine-tuned from scratch to cope with caption generation. Comparing INP with the recently proposed CLIP (Contrastive Language-Image Pre-training), this paper investigates the potential deficiencies of INP for video captioning and explores the key to generating accurate descriptions. Specifically, our empirical study on INP vs. CLIP shows that INP makes video caption models tricky to capture attributes' semantics and sensitive to irrelevant background information. By contrast, CLIP's significant boost in caption quality highlights the importance of attribute-aware representation learning. We are thus motivated to introduce Dual Attribute Prediction, an auxiliary task requiring a video caption model to learn the correspondence between video content and attributes and the co-occurrence relations between attributes. Extensive experiments on benchmark datasets demonstrate that our approach enables better learning of attribute-aware representations, bringing consistent improvements on models with different architectures and decoding algorithms.
Multi-view stereo (MVS) is a crucial task for precise 3D reconstruction. Most recent studies tried to improve the performance of matching cost volume in MVS by designing aggregated 3D cost volumes and their regularization. This paper focuses on learning a robust feature extraction network to enhance the performance of matching costs without heavy computation in the other steps. In particular, we present a dynamic scale feature extraction network, namely, CDSFNet. It is composed of multiple novel convolution layers, each of which can select a proper patch scale for each pixel guided by the normal curvature of the image surface. As a result, CDFSNet can estimate the optimal patch scales to learn discriminative features for accurate matching computation between reference and source images. By combining the robust extracted features with an appropriate cost formulation strategy, our resulting MVS architecture can estimate depth maps more precisely. Extensive experiments showed that the proposed method outperforms other state-of-the-art methods on complex outdoor scenes. It significantly improves the completeness of reconstructed models. As a result, the method can process higher resolution inputs within faster run-time and lower memory than other MVS methods. Our source code is available at url{https://github.com/TruongKhang/cds-mvsnet}.
Modern cameras are equipped with a wide array of sensors that enable recording the geospatial context of an image. Taking advantage of this, we explore depth estimation under the assumption that the camera is geocalibrated, a problem we refer to as geo-enabled depth estimation. Our key insight is that if capture location is known, the corresponding overhead viewpoint offers a valuable resource for understanding the scale of the scene. We propose an end-to-end architecture for depth estimation that uses geospatial context to infer a synthetic ground-level depth map from a co-located overhead image, then fuses it inside of an encoder/decoder style segmentation network. To support evaluation of our methods, we extend a recently released dataset with overhead imagery and corresponding height maps. Results demonstrate that integrating geospatial context significantly reduces error compared to baselines, both at close ranges and when evaluating at much larger distances than existing benchmarks consider.
Referring image segmentation aims to predict the foreground mask of the object referred by a natural language sentence. Multimodal context of the sentence is crucial to distinguish the referent from the background. Existing methods either insufficiently or redundantly model the multimodal context. To tackle this problem, we propose a "gather-propagate-distribute" scheme to model multimodal context by cross-modal interaction and implement this scheme as a novel Linguistic Structure guided Context Modeling (LSCM) module. Our LSCM module builds a Dependency Parsing Tree suppressed Word Graph (DPT-WG) which guides all the words to include valid multimodal context of the sentence while excluding disturbing ones through three steps over the multimodal feature, i.e., gathering, constrained propagation and distributing. Extensive experiments on four benchmarks demonstrate that our method outperforms all the previous state-of-the-arts.
Neural networks often make predictions relying on the spurious correlations from the datasets rather than the intrinsic properties of the task of interest, facing sharp degradation on out-of-distribution (OOD) test data. Existing de-bias learning frameworks try to capture specific dataset bias by bias annotations, they fail to handle complicated OOD scenarios. Others implicitly identify the dataset bias by the special design on the low capability biased model or the loss, but they degrade when the training and testing data are from the same distribution. In this paper, we propose a General Greedy De-bias learning framework (GGD), which greedily trains the biased models and the base model like gradient descent in functional space. It encourages the base model to focus on examples that are hard to solve with biased models, thus remaining robust against spurious correlations in the test stage. GGD largely improves models' OOD generalization ability on various tasks, but sometimes over-estimates the bias level and degrades on the in-distribution test. We further re-analyze the ensemble process of GGD and introduce the Curriculum Regularization into GGD inspired by curriculum learning, which achieves a good trade-off between in-distribution and out-of-distribution performance. Extensive experiments on image classification, adversarial question answering, and visual question answering demonstrate the effectiveness of our method. GGD can learn a more robust base model under the settings of both task-specific biased models with prior knowledge and self-ensemble biased model without prior knowledge.
Deep generative modeling has the potential to cause significant harm to society. Recognizing this threat, a magnitude of research into detecting so-called "Deepfakes" has emerged. This research most often focuses on the image domain, while studies exploring generated audio signals have, so-far, been neglected. In this paper we make three key contributions to narrow this gap. First, we provide researchers with an introduction to common signal processing techniques used for analyzing audio signals. Second, we present a novel data set, for which we collected nine sample sets from five different network architectures, spanning two languages. Finally, we supply practitioners with two baseline models, adopted from the signal processing community, to facilitate further research in this area.
In this paper, we propose a new class of high-efficient deep joint source-channel coding methods that can closely adapt to the source distribution under the nonlinear transform, it can be collected under the name nonlinear transform source-channel coding (NTSCC). In the considered model, the transmitter first learns a nonlinear analysis transform to map the source data into latent space, then transmits the latent representation to the receiver via deep joint source-channel coding. Our model incorporates the nonlinear transform as a strong prior to effectively extract the source semantic features and provide side information for source-channel coding. Unlike existing conventional deep joint source-channel coding methods, the proposed NTSCC essentially learns both the source latent representation and an entropy model as the prior on the latent representation. Accordingly, novel adaptive rate transmission and hyperprior-aided codec refinement mechanisms are developed to upgrade deep joint source-channel coding. The whole system design is formulated as an optimization problem whose goal is to minimize the end-to-end transmission rate-distortion performance under established perceptual quality metrics. Across simple example sources and test image sources, we find that the proposed NTSCC transmission method generally outperforms both the analog transmission using the standard deep joint source-channel coding and the classical separation-based digital transmission. Notably, the proposed NTSCC method can potentially support future semantic communications due to its vigorous content-aware ability.
Rate adaption of deep image compression in a single model will become one of the decisive factors competing with the classical image compression codecs. However, until now, there is no perfect solution that neither increases the computation nor affects the compression performance. In this paper, we propose a novel image compression framework G-VAE (Gained Variational Autoencoder), which could achieve continuously variable rate in a single model. Unlike the previous solutions that encode progressively or change the internal unit of the network, G-VAE only adds a pair of gain units at the output of encoder and the input of decoder. It is so concise that G-VAE could be applied to almost all the image compression methods and achieve continuously variable rate with negligible additional parameters and computation. We also propose a new deep image compression framework, which outperforms all the published results on Kodak datasets in PSNR and MS-SSIM metrics. Experimental results show that adding a pair of gain units will not affect the performance of the basic models while endowing them with continuously variable rate.
Speckle fluctuations seriously limit the interpretability of synthetic aperture radar (SAR) images. Speckle reduction has thus been the subject of numerous works spanning at least four decades. Techniques based on deep neural networks have recently achieved a new level of performance in terms of SAR image restoration quality. Beyond the design of suitable network architectures or the selection of adequate loss functions, the construction of training sets is of uttermost importance. So far, most approaches have considered a supervised training strategy: the networks are trained to produce outputs as close as possible to speckle-free reference images. Speckle-free images are generally not available, which requires resorting to natural or optical images or the selection of stable areas in long time series to circumvent the lack of ground truth. Self-supervision, on the other hand, avoids the use of speckle-free images. We introduce a self-supervised strategy based on the separation of the real and imaginary parts of single-look complex SAR images, called MERLIN (coMplex sElf-supeRvised despeckLINg), and show that it offers a straightforward way to train all kinds of deep despeckling networks. Networks trained with MERLIN take into account the spatial correlations due to the SAR transfer function specific to a given sensor and imaging mode. By requiring only a single image, and possibly exploiting large archives, MERLIN opens the door to hassle-free as well as large-scale training of despeckling networks. The code of the trained models is made freely available at https://gitlab.telecom-paris.fr/RING/MERLIN.
Recent work in deep learning focuses on solving physical systems in the Ordinary Differential Equation or Partial Differential Equation. This current work proposed a variant of Convolutional Neural Networks (CNNs) that can learn the hidden dynamics of a physical system using ordinary differential equation (ODEs) systems (ODEs) and Partial Differential Equation systems (PDEs). Instead of considering the physical system such as image, time -series as a system of multiple layers, this new technique can model a system in the form of Differential Equation (DEs). The proposed method has been assessed by solving several steady-state PDEs on irregular domains, including heat equations, Navier-Stokes equations.