We propose MONet, a convolutional neural network that jointly detects motion boundaries (MBs) and occlusion regions (Occs) in video both forward and backward in time. Detection is difficult because optical flow is discontinuous along MBs and undefined in Occs, while many flow estimators assume smoothness and a flow defined everywhere. To reason in the two time directions simultaneously, we direct-warp the estimated maps between the two frames. Since appearance mismatches between frames often signal vicinity to MBs or Occs, we construct a cost block that for each feature in one frame records the lowest discrepancy with matching features in a search range. This cost block is two-dimensional, and much less expensive than the four-dimensional cost volumes used in flow analysis. Cost-block features are computed by an encoder, and MB and Occ estimates are computed by a decoder. We found that arranging decoder layers fine-to-coarse, rather than coarse-to-fine, improves performance. MONet outperforms the prior state of the art for both tasks on the Sintel and FlyingChairsOcc benchmarks without any fine-tuning on them.
Recent years have witnessed a growing interest in research related to the detection of piano pedals from audio signals in the music information retrieval community. However, to our best knowledge, recent generative models for symbolic music have rarely taken piano pedals into account. In this work, we employ the transcription model proposed by Kong et al. to get pedal information from the audio recordings of piano performance in the AILabs1k7 dataset, and then modify the Compound Word Transformer proposed by Hsiao et al. to build a Transformer decoder that generates pedal-related tokens along with other musical tokens. While the work is done by using inferred sustain pedal information as training data, the result shows hope for further improvement and the importance of the involvement of sustain pedal in tasks of piano performance generations.
In typical multi-talker speech recognition systems, a neural network-based acoustic model predicts senone state posteriors for each speaker. These are later used by a single-talker decoder which is applied on each speaker-specific output stream separately. In this work, we argue that such a scheme is sub-optimal and propose a principled solution that decodes all speakers jointly. We modify the acoustic model to predict joint state posteriors for all speakers, enabling the network to express uncertainty about the attribution of parts of the speech signal to the speakers. We employ a joint decoder that can make use of this uncertainty together with higher-level language information. For this, we revisit decoding algorithms used in factorial generative models in early multi-talker speech recognition systems. In contrast with these early works, we replace the GMM acoustic model with DNN, which provides greater modeling power and simplifies part of the inference. We demonstrate the advantage of joint decoding in proof of concept experiments on a mixed-TIDIGITS dataset.
In this paper, we address the problem of forecasting the trajectory of an egocentric camera wearer (ego-person) in crowded spaces. The trajectory forecasting ability learned from the data of different camera wearers walking around in the real world can be transferred to assist visually impaired people in navigation, as well as to instill human navigation behaviours in mobile robots, enabling better human-robot interactions. To this end, a novel egocentric human trajectory forecasting dataset was constructed, containing real trajectories of people navigating in crowded spaces wearing a camera, as well as extracted rich contextual data. We extract and utilize three different modalities to forecast the trajectory of the camera wearer, i.e., his/her past trajectory, the past trajectories of nearby people, and the environment such as the scene semantics or the depth of the scene. A Transformer-based encoder-decoder neural network model, integrated with a novel cascaded cross-attention mechanism that fuses multiple modalities, has been designed to predict the future trajectory of the camera wearer. Extensive experiments have been conducted, with results showing that our model outperforms the state-of-the-art methods in egocentric human trajectory forecasting.
Damage assessment after natural disasters is needed to distribute aid and forces to recovery from damage dealt optimally. This process involves acquiring satellite imagery for the region of interest, localization of buildings, and classification of the amount of damage caused by nature or urban factors to buildings. In case of natural disasters, this means processing many square kilometers of the area to judge whether a particular building had suffered from the damaging factors. In this work, we develop a computational approach for an automated comparison of the same region's satellite images before and after the disaster, and classify different levels of damage in buildings. Our solution is based on Siamese neural networks with encoder-decoder architecture. We include an extensive ablation study and compare different encoders, decoders, loss functions, augmentations, and several methods to combine two images. The solution achieved one of the best results in the Computer Vision for Building Damage Assessment competition.
Predicting linearized Abstract Meaning Representation (AMR) graphs using pre-trained sequence-to-sequence Transformer models has recently led to large improvements on AMR parsing benchmarks. These parsers are simple and avoid explicit modeling of structure but lack desirable properties such as graph well-formedness guarantees or built-in graph-sentence alignments. In this work we explore the integration of general pre-trained sequence-to-sequence language models and a structure-aware transition-based approach. We depart from a pointer-based transition system and propose a simplified transition set, designed to better exploit pre-trained language models for structured fine-tuning. We also explore modeling the parser state within the pre-trained encoder-decoder architecture and different vocabulary strategies for the same purpose. We provide a detailed comparison with recent progress in AMR parsing and show that the proposed parser retains the desirable properties of previous transition-based approaches, while being simpler and reaching the new parsing state of the art for AMR 2.0, without the need for graph re-categorization.
The predominant approach for language modeling is to process sequences from left to right, but this eliminates a source of information: the order by which the sequence was generated. One strategy to recover this information is to decode both the content and ordering of tokens. Existing approaches supervise content and ordering by designing problem-specific loss functions and pre-training with an ordering pre-selected. Other recent works use iterative search to discover problem-specific orderings for training, but suffer from high time complexity and cannot be efficiently parallelized. We address these limitations with an unsupervised parallelizable learner that discovers high-quality generation orders purely from training data -- no domain knowledge required. The learner contains an encoder network and decoder language model that perform variational inference with autoregressive orders (represented as permutation matrices) as latent variables. The corresponding ELBO is not differentiable, so we develop a practical algorithm for end-to-end optimization using policy gradients. We implement the encoder as a Transformer with non-causal attention that outputs permutations in one forward pass. Permutations then serve as target generation orders for training an insertion-based Transformer language model. Empirical results in language modeling tasks demonstrate that our method is context-aware and discovers orderings that are competitive with or even better than fixed orders.
Latent dynamics models have emerged as powerful tools for modeling and interpreting neural population activity. Recently, there has been a focus on incorporating simultaneously measured behaviour into these models to further disentangle sources of neural variability in their latent space. These approaches, however, are limited in their ability to capture the underlying neural dynamics (e.g. linear) and in their ability to relate the learned dynamics back to the observed behaviour (e.g. no time lag). To this end, we introduce Targeted Neural Dynamical Modeling (TNDM), a nonlinear state-space model that jointly models the neural activity and external behavioural variables. TNDM decomposes neural dynamics into behaviourally relevant and behaviourally irrelevant dynamics; the relevant dynamics are used to reconstruct the behaviour through a flexible linear decoder and both sets of dynamics are used to reconstruct the neural activity through a linear decoder with no time lag. We implement TNDM as a sequential variational autoencoder and validate it on simulated recordings and recordings taken from the premotor and motor cortex of a monkey performing a center-out reaching task. We show that TNDM is able to learn low-dimensional latent dynamics that are highly predictive of behaviour without sacrificing its fit to the neural data.
Feature fusion and similarity computation are two core problems in 3D object tracking, especially for object tracking using sparse and disordered point clouds. Feature fusion could make similarity computing more efficient by including target object information. However, most existing LiDAR-based approaches directly use the extracted point cloud feature to compute similarity while ignoring the attention changes of object regions during tracking. In this paper, we propose a feature fusion network based on transformer architecture. Benefiting from the self-attention mechanism, the transformer encoder captures the inter- and intra- relations among different regions of the point cloud. By using cross-attention, the transformer decoder fuses features and includes more target cues into the current point cloud feature to compute the region attentions, which makes the similarity computing more efficient. Based on this feature fusion network, we propose an end-to-end point cloud object tracking framework, a simple yet effective method for 3D object tracking using point clouds. Comprehensive experimental results on the KITTI dataset show that our method achieves new state-of-the-art performance. Code is available at: https://github.com/3bobo/lttr.
We introduce Imuge, an image tamper resilient generative scheme for image self-recovery. The traditional manner of concealing image content within the image are inflexible and fragile to diverse digital attack, i.e. image cropping and JPEG compression. To address this issue, we jointly train a U-Net backboned encoder, a tamper localization network and a decoder for image recovery. Given an original image, the encoder produces a visually indistinguishable immunized image. At the recipient's side, the verifying network localizes the malicious modifications, and the original content can be approximately recovered by the decoder, despite the presence of the attacks. Several strategies are proposed to boost the training efficiency. We demonstrate that our method can recover the details of the tampered regions with a high quality despite the presence of various kinds of attacks. Comprehensive ablation studies are conducted to validate our network designs.
At the heart of the success of deep learning is the quality of the data. Through data augmentation, one can train models with better generalization capabilities and thus achieve greater results in their field of interest. In this work, we explore how to augment a varied set of image datasets through relighting so as to improve the ability of existing models to be invariant to illumination changes, namely for learned descriptors. We develop a tool, based on an encoder-decoder network, that is able to quickly generate multiple variations of the illumination of various input scenes whilst also allowing the user to define parameters such as the angle of incidence and intensity. We demonstrate that by training models on datasets that have been augmented with our pipeline, it is possible to achieve higher performance on localization benchmarks.
High Resolution (HR) medical images provide rich anatomical structure details to facilitate early and accurate diagnosis. In MRI, restricted by hardware capacity, scan time, and patient cooperation ability, isotropic 3D HR image acquisition typically requests long scan time and, results in small spatial coverage and low SNR. Recent studies showed that, with deep convolutional neural networks, isotropic HR MR images could be recovered from low-resolution (LR) input via single image super-resolution (SISR) algorithms. However, most existing SISR methods tend to approach a scale-specific projection between LR and HR images, thus these methods can only deal with a fixed up-sampling rate. For achieving different up-sampling rates, multiple SR networks have to be built up respectively, which is very time-consuming and resource-intensive. In this paper, we propose ArSSR, an Arbitrary Scale Super-Resolution approach for recovering 3D HR MR images. In the ArSSR model, the reconstruction of HR images with different up-scaling rates is defined as learning a continuous implicit voxel function from the observed LR images. Then the SR task is converted to represent the implicit voxel function via deep neural networks from a set of paired HR-LR training examples. The ArSSR model consists of an encoder network and a decoder network. Specifically, the convolutional encoder network is to extract feature maps from the LR input images and the fully-connected decoder network is to approximate the implicit voxel function. Due to the continuity of the learned function, a single ArSSR model can achieve arbitrary up-sampling rate reconstruction of HR images from any input LR image after training. Experimental results on three datasets show that the ArSSR model can achieve state-of-the-art SR performance for 3D HR MR image reconstruction while using a single trained model to achieve arbitrary up-sampling scales.
As one of the most popular generative models, Variational Autoencoder (VAE) approximates the posterior of latent variables based on amortized variational inference. However, when the decoder network is sufficiently expressive, VAE may lead to posterior collapse; that is, uninformative latent representations may be learned. To this end, in this paper, we propose an alternative model, DU-VAE, for learning a more Diverse and less Uncertain latent space, and thus the representation can be learned in a meaningful and compact manner. Specifically, we first theoretically demonstrate that it will result in better latent space with high diversity and low uncertainty awareness by controlling the distribution of posterior's parameters across the whole data accordingly. Then, without the introduction of new loss terms or modifying training strategies, we propose to exploit Dropout on the variances and Batch-Normalization on the means simultaneously to regularize their distributions implicitly. Furthermore, to evaluate the generalization effect, we also exploit DU-VAE for inverse autoregressive flow based-VAE (VAE-IAF) empirically. Finally, extensive experiments on three benchmark datasets clearly show that our approach can outperform state-of-the-art baselines on both likelihood estimation and underlying classification tasks.
Semi-supervised learning (SSL) is a promising machine learning paradigm to address the issue of label scarcity in medical imaging. SSL methods were originally developed in image classification. The state-of-the-art SSL methods in image classification utilise consistency regularisation to learn unlabelled predictions which are invariant to input level perturbations. However, image level perturbations violate the cluster assumption in the setting of segmentation. Moreover, existing image level perturbations are hand-crafted which could be sub-optimal. Therefore, it is a not trivial to straightforwardly adapt existing SSL image classification methods in segmentation. In this paper, we propose MisMatch, a semi-supervised segmentation framework based on the consistency between paired predictions which are derived from two differently learnt morphological feature perturbations. MisMatch consists of an encoder and two decoders. One decoder learns positive attention for foreground on unlabelled data thereby generating dilated features of foreground. The other decoder learns negative attention for foreground on the same unlabelled data thereby generating eroded features of foreground. We first develop a 2D U-net based MisMatch framework and perform extensive cross-validation on a CT-based pulmonary vessel segmentation task and show that MisMatch statistically outperforms state-of-the-art semi-supervised methods when only 6.25\% of the total labels are used. In a second experiment, we show that U-net based MisMatch outperforms state-of-the-art methods on an MRI-based brain tumour segmentation task. In a third experiment, we show that a 3D MisMatch outperforms a previous method using input level augmentations, on a left atrium segmentation task. Lastly, we find that the performance improvement of MisMatch over the baseline might originate from its better calibration.
Automatic Image Captioning is the never-ending effort of creating syntactically and validating the accuracy of textual descriptions of an image in natural language with context. The encoder-decoder structure used throughout existing Bengali Image Captioning (BIC) research utilized abstract image feature vectors as the encoder's input. We propose a novel transformer-based architecture with an attention mechanism with a pre-trained ResNet-101 model image encoder for feature extraction from images. Experiments demonstrate that the language decoder in our technique captures fine-grained information in the caption and, then paired with image features, produces accurate and diverse captions on the BanglaLekhaImageCaptions dataset. Our approach outperforms all existing Bengali Image Captioning work and sets a new benchmark by scoring 0.694 on BLEU-1, 0.630 on BLEU-2, 0.582 on BLEU-3, and 0.337 on METEOR.
Clothes style transfer for person video generation is a challenging task, due to drastic variations of intra-person appearance and video scenarios. To tackle this problem, most recent AdaIN-based architectures are proposed to extract clothes and scenario features for generation. However, these approaches suffer from being short of fine-grained details and are prone to distort the origin person. To further improve the generation performance, we propose a novel framework with disentangled multi-branch encoders and a shared decoder. Moreover, to pursue the strong video spatio-temporal consistency, an inner-frame discriminator is delicately designed with input being cross-frame difference. Besides, the proposed framework possesses the property of scenario adaptation. Extensive experiments on the TEDXPeople benchmark demonstrate the superiority of our method over state-of-the-art approaches in terms of image quality and video coherence.
Although DALL-E has shown an impressive ability of composition-based systematic generalization in image generation, it requires the dataset of text-image pairs and the compositionality is provided by the text. In contrast, object-centric representation models like the Slot Attention model learn composable representations without the text prompt. However, unlike DALL-E its ability to systematically generalize for zero-shot generation is significantly limited. In this paper, we propose a simple but novel slot-based autoencoding architecture, called SLATE, for combining the best of both worlds: learning object-centric representations that allows systematic generalization in zero-shot image generation without text. As such, this model can also be seen as an illiterate DALL-E model. Unlike the pixel-mixture decoders of existing object-centric representation models, we propose to use the Image GPT decoder conditioned on the slots for capturing complex interactions among the slots and pixels. In experiments, we show that this simple and easy-to-implement architecture not requiring a text prompt achieves significant improvement in in-distribution and out-of-distribution (zero-shot) image generation and qualitatively comparable or better slot-attention structure than the models based on mixture decoders.
This paper introduces Attentive Implicit Representation Networks (AIR-Nets), a simple, but highly effective architecture for 3D reconstruction from point clouds. Since representing 3D shapes in a local and modular fashion increases generalization and reconstruction quality, AIR-Nets encode an input point cloud into a set of local latent vectors anchored in 3D space, which locally describe the object's geometry, as well as a global latent description, enforcing global consistency. Our model is the first grid-free, encoder-based approach that locally describes an implicit function. The vector attention mechanism from [Zhao et al. 2020] serves as main point cloud processing module, and allows for permutation invariance and translation equivariance. When queried with a 3D coordinate, our decoder gathers information from the global and nearby local latent vectors in order to predict an occupancy value. Experiments on the ShapeNet dataset show that AIR-Nets significantly outperform previous state-of-the-art encoder-based, implicit shape learning methods and especially dominate in the sparse setting. Furthermore, our model generalizes well to the FAUST dataset in a zero-shot setting. Finally, since AIR-Nets use a sparse latent representation and follow a simple operating scheme, the model offers several exiting avenues for future work. Our code is available at https://github.com/SimonGiebenhain/AIR-Nets.
Clustering is an unsupervised data mining technique that can be employed to segment customers. The efficient clustering of customers enables banks to design and make offers based on the features of the target customers. The present study uses a real-world financial dataset (Berka, 2000) to cluster bank customers by an encoder-decoder network and the dynamic time warping (DTW) method. The customer features required for clustering are obtained in four ways: Dynamic Time Warping (DTW), Recency Frequency and Monetary (RFM), LSTM encoder-decoder network, and our proposed hybrid method. Once the LSTM model was trained by customer transaction data, a feature vector of each customer was automatically extracted by the encoder.Moreover, the distance between pairs of sequences of transaction amounts was obtained using DTW. Another vector feature was calculated for customers by RFM scoring. In the hybrid method, the feature vectors are combined from the encoder-decoder output, the DTW distance, and the demographic data (e.g., age and gender). Finally, feature vectors were introduced as input to the k-means clustering algorithm, and we compared clustering results with Silhouette and Davies-Bouldin index. As a result, the clusters obtained from the hybrid approach are more accurate and meaningful than those derived from individual clustering techniques. In addition, the type of neural network layers had a substantial effect on the clusters, and high network error does not necessarily worsen clustering performance.
It is hard to generate an image at target view well for previous cross-view image translation methods that directly adopt a simple encoder-decoder or U-Net structure, especially for drastically different views and severe deformation cases. To ease this problem, we propose a novel two-stage framework with a new Cascaded Cross MLP-Mixer (CrossMLP) sub-network in the first stage and one refined pixel-level loss in the second stage. In the first stage, the CrossMLP sub-network learns the latent transformation cues between image code and semantic map code via our novel CrossMLP blocks. Then the coarse results are generated progressively under the guidance of those cues. Moreover, in the second stage, we design a refined pixel-level loss that eases the noisy semantic label problem with more reasonable regularization in a more compact fashion for better optimization. Extensive experimental results on Dayton~\cite{vo2016localizing} and CVUSA~\cite{workman2015wide} datasets show that our method can generate significantly better results than state-of-the-art methods. The source code and trained models are available at https://github.com/Amazingren/CrossMLP.