AAAI 2021 论文列表

Although the recent image-based 3D object detection methods using Pseudo-LiDAR representation have shown great capabilities, a notable gap in efficiency and accuracy still exist compared with LiDAR-based methods. Besides, over-reliance on the stand-alone depth estimator, requiring a large number of pixel-wise annotations in the training stage and more computation in the inferencing stage, limits the scaling application in the real world. In this paper, we propose an efficient and accurate 3D object detection method from stereo images, named RTS3D. Different from the 3D occupancy space in the Pseudo-LiDAR similar methods, we design a novel 4D feature-consistent embedding (FCE) space as the intermediate representation of the 3D scene without depth supervision. The FCE space encodes the object's structural and semantic information by exploring the multi-scale feature consistency warped from stereo pair. Furthermore, a semantic-guided RBF (Radial Basis Function) and a structure-aware attention module are devised to reduce the influence of FCE space noise without instance mask supervision. Experiments on KITTI benchmark show that RTS3D is the first true real-time system (FPS>24) for stereo image 3D detection meanwhile achieves 10% improvement in average precision comparing with the previous state-of-the-art method.

Mobile devices are becoming an important carrier for deep learning tasks, as they are being equipped with powerful, high-end mobile CPUs and GPUs. However, it is still a challenging task to execute 3D Convolutional Neural Networks (CNNs) targeting for real-time performance, besides high inference accuracy. The reason is more complex model structure and higher model dimensionality overwhelm the available computation/storage resources on mobile devices. A natural way may be turning to deep learning weight pruning techniques. However, the direct generalization of existing 2D CNN weight pruning methods to 3D CNNs is not ideal for fully exploiting mobile parallelism while achieving high inference accuracy. This paper proposes RT3D, a model compression and mobile acceleration framework for 3D CNNs, seamlessly integrating neural network weight pruning and compiler code generation techniques. We propose and investigate two structured sparsity schemes i.e., the vanilla structured sparsity and kernel group structured (KGS) sparsity that are mobile acceleration friendly. The vanilla sparsity removes whole kernel groups, while KGS sparsity is a more fine-grained structured sparsity that enjoys higher flexibility while exploiting full on-device parallelism. We propose a reweighted regularization pruning algorithm to achieve the proposed sparsity schemes. The inference time speedup due to sparsity is approaching the pruning rate of the whole model FLOPs (floating point operations). RT3D demonstrates up to 29.1x speedup in end-to-end inference time comparing with current mobile frameworks supporting 3D CNNs, with moderate 1%~1.5% accuracy loss. The end-to-end inference time for 16 video frames could be within 150 ms, when executing representative C3D and R(2+1)D models on a cellphone. For the first time, real-time execution of 3D CNNs is achieved on off-the-shelf mobiles.

We study unsupervised video representation learning that seeks to learn both motion and appearance features from unlabeled video only, which can be reused for downstream tasks such as action recognition. This task, however, is extremely challenging due to 1) the highly complex spatial-temporal information in videos and 2) the lack of labeled data for training. Unlike representation learning for static images, it is difficult to construct a suitable self-supervised task to effectively model both motion and appearance features. More recently, several attempts have been made to learn video representation through video playback speed prediction. However, it is non-trivial to obtain precise speed labels for the videos. More critically, the learned models may tend to focus on motion patterns and thus may not learn appearance features well. In this paper, we observe that the relative playback speed is more consistent with motion patterns and thus provides more effective and stable supervision for representation learning. Therefore, we propose a new way to perceive the playback speed and exploit the relative speed between two video clips as labels. In this way, we are able to effectively perceive speed and learn better motion features. Moreover, to ensure the learning of appearance features, we further propose an appearance-focused task, where we enforce the model to perceive the appearance difference between two video clips. We show that jointly optimizing the two tasks consistently improves the performance on two downstream tasks (namely, action recognition and video retrieval) w.r.t the increasing pre-training epochs. Remarkably, for action recognition on the UCF101 dataset, we achieve 93.7% accuracy without the use of labeled data for pre-training, which outperforms the ImageNet supervised pre-trained model. Our code, pre-trained models, and supplementary materials can be found at https://github.com/PeihaoChen/RSPNet.

Despite of the recent great progress on multi-person pose estimation, existing solutions still remain challenging under the condition of "crowded scenes'', where RGB images capture complex real-world scenes with highly-overlapped people, severe occlusions and diverse postures. In this work, we focus on two main problems: 1) how to design an effective pipeline for crowded scenes pose estimation; and 2) how to equip this pipeline with the ability of relation modeling for interference resolving. To tackle these problems, we propose a new pipeline named Relation based Skeleton Graph Network (RSGNet). Unlike existing works that directly predict joints-of-target by labeling joints-of-interference as false positive, we first encourage all joints to be predicted. And then, a Target-aware Relation Parser (TRP) is designed to model the relation over all predicted joints, resulting in a target-aware encoding. This new pipeline will largely relieve the confusion of the joints estimation model when seeing identical joints with totally distinct labels (e.g., the identical hand exists in two bounding boxes). Furthermore, we introduce a Skeleton Graph Machine (SGM) to model the skeleton-based commonsense knowledge, aiming to estimate the target pose with the constraint of human body structure. Such skeleton-based constraint can help to deal with the challenges in crowded scenes from a reasoning perspective. Solid experiments on pose estimation benchmarks demonstrate that our method outperforms existing state-of-the-art methods.

Pre-trained language models of the BERT family have defined the state-of-the-arts in a wide range of NLP tasks. However, the performance of BERT-based models is mainly driven by the enormous amount of parameters, which hinders their application to resource-limited scenarios. Faced with this problem, recent studies have been attempting to compress BERT into a small-scale model. However, most previous work primarily focuses on a single kind of compression technique, and few attention has been paid to the combination of different methods. When BERT is compressed with integrated techniques, a critical question is how to design the entire compression framework to obtain the optimal performance. In response to this question, we integrate three kinds of compression methods (weight pruning, low-rank factorization and knowledge distillation (KD)) and explore a range of designs concerning model architecture, KD strategy, pruning frequency and learning rate schedule. We find that a careful choice of the designs is crucial to the performance of the compressed model. Based on the empirical findings, our best compressed model, dubbed Refined BERT cOmpreSsion with InTegrAted techniques (ROSITA), is 7.5x smaller than BERT while maintains 98.5% of the performance on five tasks of the GLUE benchmark, outperforming the previous BERT compression methods with similar parameter budget.

RNA-binding proteins (RBPs) play a significant part in several biological processes in the living cell, such as gene regulation and mRNA localization. Several deep learning methods, especially the model based on convolutional neural network(CNN), have been used to predict the binding sites. However, previous methods fail to represent RNA secondary structure features. The traditional deep learning methods generally transform the RNA secondary structure to a regular matrix that cannot reveal the topological structure information of RNA. To effectively extract the structure features of RNA, we propose an RNA secondary structure representation network (RNASSR-Net) based on graph convolutional neural network (GCN) and convolution neural network (CNN) for RBP binding prediction. RNASSR-Net constructs the graph model derived from the RNA secondary structure to learn the topological properties of RNA. Then, it obtains the spatial importance of each base in RNA with CNN to guide the representation of the RNA secondary structure. Finally, RNASSR-Net combines the structure and sequence features to predict the binding sites. Experimental results demonstrate the proposed method outperforms a few state-of-the-art methods on the benchmark datasets and gets a higher improvement on the small-size data. Besides, the proposed RNASSR-Net is also used to detect the accurate motifs compared with the experimentally verified motifs, which reveals the binding region location and RNA structure interpretation for some biological guidance in the future.

RGB-D salient object detection (SOD) recently has attracted increasing research interest and many deep learning methods based on encoder-decoder architectures have emerged. However, most existing RGB-D SOD models conduct feature fusion either in the single encoder or the decoder stage, which hardly guarantees sufficient cross-modal fusion ability. In this paper, we make the first attempt in addressing RGB-D SOD through 3D convolutional neural networks. The proposed model, named RD3D, aims at pre-fusion in the encoder stage and in-depth fusion in the decoder stage to effectively promote the full integration of RGB and depth streams. Specifically, RD3D first conducts pre-fusion across RGB and depth modalities through an inflated 3D encoder, and later provides in-depth feature fusion by designing a 3D decoder equipped with rich back-projection paths (RBPP) for leveraging the extensive aggregation ability of 3D convolutions. With such a progressive fusion strategy involving both the encoder and decoder, effective and thorough interaction between the two modalities can be exploited and boost the detection accuracy. Extensive experiments on six widely used benchmark datasets demonstrate that RD3D performs favorably against 14 state-of-the-art RGB-D SOD approaches in terms of four key evaluation metrics. Our code will be made publicly available: https://github.com/PPOLYpubki/RD3D.

Lane detection is one of the most important tasks in self-driving. Due to various complex scenarios (e.g., severe occlusion, ambiguous lanes, etc.) and the sparse supervisory signals inherent in lane annotations, lane detection task is still challenging. Thus, it is difficult for the ordinary convolutional neural network (CNN) to train in general scenes to catch subtle lane feature from the raw image. In this paper, we present a novel module named REcurrent Feature-Shift Aggregator (RESA) to enrich lane feature after preliminary feature extraction with an ordinary CNN. RESA takes advantage of strong shape priors of lanes and captures spatial relationships of pixels across rows and columns. It shifts sliced feature map recurrently in vertical and horizontal directions and enables each pixel to gather global information. RESA can conjecture lanes accurately in challenging scenarios with weak appearance clues by aggregating sliced feature map. Moreover, we propose a Bilateral Up-Sampling Decoder that combines coarse-grained and fine-detailed features in the up-sampling stage. It can recover the low-resolution feature map into pixel-wise prediction meticulously. Our method achieves state-of-the-art results on two popular lane detection benchmarks (CULane and Tusimple). Code has been made available at: https://github.com/ZJULearning/resa.

When answering a question, people often draw upon their rich world knowledge in addition to the particular context. While recent works retrieve supporting facts/evidence from commonsense knowledge bases to supply additional information to each question, there is still ample opportunity to advance it on the quality of the evidence. It is crucial since the quality of the evidence is the key to answering common- sense questions, and even determines the upper bound on the QA systems’ performance. In this paper, we propose a recursive erasure memory network (REM-Net) to cope with the quality improvement of evidence. To address this, REM-Net is equipped with a module to refine the evidence by recursively erasing the low-quality evidence that does not explain the question answering. Besides, instead of retrieving evidence from existing knowledge bases, REM-Net leverages a pre-trained generative model to generate candidate evidence customized for the question. We conduct experiments on two commonsense question answering datasets, WIQA and CosmosQA. The results demonstrate the performance of REM- Net and show that the refined evidence is explainable.

Panoptic segmentation aims at generating pixel-wise class and instance predictions for each pixel in the input image, which is a challenging task and far more complicated than naively fusing the semantic and instance segmentation results. Prediction fusion is therefore important to achieve accurate panoptic segmentation. In this paper, we present REFINE, pREdiction FusIon NEtwork for panoptic segmentation, to achieve high-quality panoptic segmentation by improving cross-task prediction fusion, and within-task prediction fusion. Our single-model ResNeXt-101 with DCN achieves PQ=51.5 on the COCO dataset, surpassing state-of-the-art performance by a convincing margin and is comparable with ensembled models. Our smaller model with a ResNet-50 backbone achieves PQ=44.9, which is comparable with state-of-the-art methods with larger backbones.

Rotation detection is a challenging task due to the difficulties of locating the multi-angle objects and separating them effectively from the background. Though considerable progress has been made, for practical settings, there still exist challenges for rotating objects with large aspect ratio, dense distribution and category extremely imbalance. In this paper, we propose an end-to-end refined single-stage rotation detector for fast and accurate object detection by using a progressive regression approach from coarse to fine granularity. Considering the shortcoming of feature misalignment in existing refined single-stage detector, we design a feature refinement module to improve detection performance by getting more accurate features. The key idea of feature refinement module is to re-encode the position information of the current refined bounding box to the corresponding feature points through pixel-wise feature interpolation to realize feature reconstruction and alignment. For more accurate rotation estimation, an approximate SkewIoU loss is proposed to solve the problem that the calculation of SkewIoU is not derivable. Experiments on three popular remote sensing public datasets DOTA, HRSC2016, UCAS-AOD as well as one scene text dataset ICDAR2015 show the effectiveness of our approach. The source code is available at https://github.com/Thinklab-SJTU/R3Det_Tensorflow and is also integrated in our open source rotation detection benchmark: https://github.com/yangxue0827/RotationDetection.

End-to-end delay is a critical attribute of quality of service (QoS) in application domains such as cloud computing and computer networks. This metric is particularly important in tandem service systems, where the end-to-end service is provided through a chain of services. Service-rate control is a common mechanism for providing QoS guarantees in service systems. In this paper, we introduce a reinforcement learning-based (RL-based) service-rate controller that provides probabilistic upper-bounds on the end-to-end delay of the system, while preventing the overuse of service resources. In order to have a general framework, we use queueing theory to model the service systems. However, we adopt an RL-based approach to avoid the limitations of queueing-theoretic methods. In particular, we use Deep Deterministic Policy Gradient (DDPG) to learn the service rates (action) as a function of the queue lengths (state) in tandem service systems. In contrast to existing RL-based methods that quantify their performance by the achieved overall reward, which could be hard to interpret or even misleading, our proposed controller provides explicit probabilistic guarantees on the end-to-end delay of the system. The evaluations are presented for a tandem queueing system with non-exponential inter-arrival and service times, the results of which validate our controller's capability in meeting QoS constraints.

Aspect term extraction and opinion word extraction are two fundamental subtasks of aspect-based sentiment analysis. The internal relationship between aspect terms and opinion words is typically ignored, and information for the decision-making of buyers and sellers is insufficient. In this paper, we explore an aspect–opinion pair extraction (AOPE) task and propose a Question-Driven Span Labeling (QDSL) model to extract all the aspect–opinion pairs from user-generated reviews. Specifically, we divide the AOPE task into aspect term extraction (ATE) and aspect-specified opinion extraction (ASOE) subtasks; we first extract all the candidate aspect terms and then the corresponding opinion words given the aspect term. Unlike existing approaches that use the BIO-based tagging scheme for extraction, the QDSL model adopts a span-based tagging scheme and builds a question–answer-based machine-reading comprehension task for an effective aspect–opinion pair extraction. Extensive experiments conducted on three tasks (ATE, ASOE, and AOPE) on four benchmark datasets demonstrate that the proposed method significantly outperforms state-of-the-art approaches.

Weakly-supervised video object segmentation (WVOS) is an emerging video task that can track and segment the target given a simple bounding box label. However, existing WVOS methods are still unsatisfied in either speed or accuracy, since they only use the exemplar frame to guide the prediction while they neglect the reference from other frames. To solve the problem, we propose a novel Re-Aggregation based framework, which uses feature matching to efficiently find the target and capture the temporal dependencies from multiple frames to guide the segmentation. Based on a two-stage structure, our framework builds an information-symmetric matching process to achieve robust aggregation. In each stage, we design a Query-Memory Aggregation (QMA) module to gather features from the past frames and make bidirectional aggregation to adaptively weight the aggregated features, which relieves the latent misguidance in unidirectional aggregation. To further exploit the information from different aggregation stages, we propose a novel coarse-fine constraint by using the Cascaded Refinement Module (CRM) to combine the predictions from different stages and further boosts the performance. Experimental results on three benchmarks show that our method achieves the state-of-the-art performance in WVOS (e.g., an overall score of 84.7% on the DAVIS 2016 validation set).

Probabilistic graphical models (PGMs) provide a compact representation of knowledge that can be queried in a flexible way: after learning the parameters of a graphical model once, new probabilistic queries can be answered at test time without retraining. However, when using undirected PGMS with hidden variables, two sources of error typically compound in all but the simplest models (a) learning error (both computing the partition function and integrating out the hidden variables is intractable); and (b) prediction error (exact inference is also intractable). Here we introduce query training (QT), a mechanism to learn a PGM that is optimized for the approximate inference algorithm that will be paired with it. The resulting PGM is a worse model of the data (as measured by the likelihood), but it is tuned to produce better marginals for a given inference algorithm. Unlike prior works, our approach preserves the querying flexibility of the original PGM: at test time, we can estimate the marginal of any variable given any partial evidence. We demonstrate experimentally that QT can be used to learn a challenging 8-connected grid Markov random field with hidden variables and that it consistently outperforms the state-of-the-art AdVIL when tested on three undirected models across multiple datasets.

We provide a novel perspective on conversational emotion recognition by drawing an analogy between the task and a complete span of quantum measurement. We characterize different steps of quantum measurement in the process of recognizing speakers' emotions in conversation, and stitch them up with a quantum-like neural network. The quantum-like layers are implemented by complex-valued operations to ensure an authentic adoption of quantum concepts, which naturally enables conversational context modeling and multimodal fusion. We borrow an existing algorithm to learn the complex-valued network weights, so that the quantum-like procedure is conducted in a data-driven manner. Our model is comparable to state-of-the-art approaches on two benchmarking datasets, and provide a quantum view to understand conversational emotion recognition.

Identifying the best arm of a multi-armed bandit is a central problem in bandit optimization. We study a quantum computational version of this problem with coherent oracle access to states encoding the reward probabilities of each arm as quantum amplitudes. Specifically, we provide an algorithm to find the best arm with fixed confidence based on variable-time amplitude amplification and estimation. This algorithm gives a quadratic speedup compared to the best possible classical result in terms of query complexity. We also prove a matching quantum lower bound (up to poly-logarithmic factors).

Video sentiment analysis as a decision-making process is inherently complex, involving the fusion of decisions from multiple modalities and the so-caused cognitive biases. Inspired by recent advances in quantum cognition, we show that the sentiment judgment from one modality could be incompatible with the judgment from another, i.e., the order matters and they cannot be jointly measured to produce a final decision. Thus the cognitive process exhibits ``quantum-like'' biases that cannot be captured by classical probability theories. Accordingly, we propose a fundamentally new, quantum cognitively motivated fusion strategy for predicting sentiment judgments. In particular, we formulate utterances as quantum superposition states of positive and negative sentiment judgments, and uni-modal classifiers as mutually incompatible observables, on a complex-valued Hilbert space with positive-operator valued measures. Experiments on two benchmarking datasets illustrate that our model significantly outperforms various existing decision level and a range of state-of-the-art content-level fusion approaches. The results also show that the concept of incompatibility allows effective handling of all combination patterns, including those extreme cases that are wrongly predicted by all uni-modal classifiers.

In a situation where an agent has to produce specific resources using the available ones, it may not be possible to achieve the complete goal, but only obtaining some of its parts.This incompleteness problem calls for reasoning models to make rational decisions. In this paper, we introduce a logic-based framework for measuring resource production incompleteness: the greater the value returned by a measure, the greater is the intensity of incompleteness. After motivating our work by describing situations where the incompleteness measures can be applied, we introduce our framework by using a postulate-based approach. To some extent, the incompleteness measures can be seen as a counterpart of inconsistency measures in resource logics. Here, intuitionistic affine logic is used for representing and reasoning about resource consummation and production. Besides, we propose different notions that are useful for defining different types of incompleteness measures. We also present several measures to illustrate the introduced concepts and notions.