Designing an efficient yet deployment-friendly 3D backbone to handle sparse point clouds is a fundamental problem in 3D object detection. Compared with the customized sparse convolution, the attention mechanism in Transformers is more appropriate for flexibly modeling long-range relationships and is easier to be deployed in real-world applications. However, due to the sparse characteristics of point clouds, it is non-trivial to apply a standard transformer on sparse points. In this paper, we present Dynamic Sparse Voxel Transformer (DSVT), a single-stride window-based voxel Transformer backbone for outdoor 3D object detection. In order to efficiently process sparse points in parallel, we propose Dynamic Sparse Window Attention, which partitions a series of local regions in each window according to its sparsity and then computes the features of all regions in a fully parallel manner. To allow the cross-set connection, we design a rotated set partitioning strategy that alternates between two partitioning configurations in consecutive self-attention layers. To support effective downsampling and better encode geometric information, we also propose an attention-style 3D pooling module on sparse points, which is powerful and deployment-friendly without utilizing any customized CUDA operations. Our model achieves state-of-the-art performance on large-scale Waymo Open Dataset with remarkable gains. More importantly, DSVT can be easily deployed by TensorRT with real-time inference speed (27Hz). Code will be available at \url{https://github.com/Haiyang-W/DSVT}.
In knowledge distillation, previous feature distillation methods mainly focus on the design of loss functions and the selection of the distilled layers, while the effect of the feature projector between the student and the teacher remains under-explored. In this paper, we first discuss a plausible mechanism of the projector with empirical evidence and then propose a new feature distillation method based on a projector ensemble for further performance improvement. We observe that the student network benefits from a projector even if the feature dimensions of the student and the teacher are the same. Training a student backbone without a projector can be considered as a multi-task learning process, namely achieving discriminative feature extraction for classification and feature matching between the student and the teacher for distillation at the same time. We hypothesize and empirically verify that without a projector, the student network tends to overfit the teacher's feature distributions despite having different architecture and weights initialization. This leads to degradation on the quality of the student's deep features that are eventually used in classification. Adding a projector, on the other hand, disentangles the two learning tasks and helps the student network to focus better on the main feature extraction task while still being able to utilize teacher features as a guidance through the projector. Motivated by the positive effect of the projector in feature distillation, we propose an ensemble of projectors to further improve the quality of student features. Experimental results on different datasets with a series of teacher-student pairs illustrate the effectiveness of the proposed method.
Deep Reinforcement Learning (DRL) has achieved impressive performance in robotics and autonomous systems (RASs). A key impediment to its deployment in real-life operations is the spuriously unsafe DRL policies--unexplored states may lead the agent to make wrong decisions that may cause hazards, especially in applications where end-to-end controllers of the RAS were trained by DRL. In this paper, we propose a novel quantitative reliability assessment framework for DRL-controlled RASs, leveraging verification evidence generated from formal reliability analysis of neural networks. A two-level verification framework is introduced to check the safety property with respect to inaccurate observations that are due to, e.g., environmental noises and state changes. Reachability verification tools are leveraged at the local level to generate safety evidence of trajectories, while at the global level, we quantify the overall reliability as an aggregated metric of local safety evidence, according to an operational profile. The effectiveness of the proposed verification framework is demonstrated and validated via experiments on real RASs.
Zero-shot learning is a learning regime that recognizes unseen classes by generalizing the visual-semantic relationship learned from the seen classes. To obtain an effective ZSL model, one may resort to curating training samples from multiple sources, which may inevitably raise the privacy concerns about data sharing across different organizations. In this paper, we propose a novel Federated Zero-Shot Learning FedZSL framework, which learns a central model from the decentralized data residing on edge devices. To better generalize to previously unseen classes, FedZSL allows the training data on each device sampled from the non-overlapping classes, which are far from the i.i.d. that traditional federated learning commonly assumes. We identify two key challenges in our FedZSL protocol: 1) the trained models are prone to be biased to the locally observed classes, thus failing to generalize to the unseen classes and/or seen classes appeared on other devices; 2) as each category in the training data comes from a single source, the central model is highly vulnerable to model replacement (backdoor) attacks. To address these issues, we propose three local objectives for visual-semantic alignment and cross-device alignment through relation distillation, which leverages the normalized class-wise covariance to regularize the consistency of the prediction logits across devices. To defend against the backdoor attacks, a feature magnitude defending technique is proposed. As malicious samples are less correlated to the given semantic attributes, the visual features of low magnitude will be discarded to stabilize model updates. The effectiveness and robustness of FedZSL are demonstrated by extensive experiments conducted on three zero-shot benchmark datasets.
Night-Time Scene Parsing (NTSP) is essential to many vision applications, especially for autonomous driving. Most of the existing methods are proposed for day-time scene parsing. They rely on modeling pixel intensity-based spatial contextual cues under even illumination. Hence, these methods do not perform well in night-time scenes as such spatial contextual cues are buried in the over-/under-exposed regions in night-time scenes. In this paper, we first conduct an image frequency-based statistical experiment to interpret the day-time and night-time scene discrepancies. We find that image frequency distributions differ significantly between day-time and night-time scenes, and understanding such frequency distributions is critical to NTSP problem. Based on this, we propose to exploit the image frequency distributions for night-time scene parsing. First, we propose a Learnable Frequency Encoder (LFE) to model the relationship between different frequency coefficients to measure all frequency components dynamically. Second, we propose a Spatial Frequency Fusion module (SFF) that fuses both spatial and frequency information to guide the extraction of spatial context features. Extensive experiments show that our method performs favorably against the state-of-the-art methods on the NightCity, NightCity+ and BDD100K-night datasets. In addition, we demonstrate that our method can be applied to existing day-time scene parsing methods and boost their performance on night-time scenes.
Deep neural networks (DNNs) are vulnerable to adversarial examples, in which DNNs are misled to false outputs due to inputs containing imperceptible perturbations. Adversarial training, a reliable and effective method of defense, may significantly reduce the vulnerability of neural networks and becomes the de facto standard for robust learning. While many recent works practice the data-centric philosophy, such as how to generate better adversarial examples or use generative models to produce additional training data, we look back to the models themselves and revisit the adversarial robustness from the perspective of deep feature distribution as an insightful complementarity. In this paper, we propose Branch Orthogonality adveRsarial Training (BORT) to obtain state-of-the-art performance with solely the original dataset for adversarial training. To practice our design idea of integrating multiple orthogonal solution spaces, we leverage a simple and straightforward multi-branch neural network that eclipses adversarial attacks with no increase in inference time. We heuristically propose a corresponding loss function, branch-orthogonal loss, to make each solution space of the multi-branch model orthogonal. We evaluate our approach on CIFAR-10, CIFAR-100, and SVHN against \ell_{\infty} norm-bounded perturbations of size \epsilon = 8/255, respectively. Exhaustive experiments are conducted to show that our method goes beyond all state-of-the-art methods without any tricks. Compared to all methods that do not use additional data for training, our models achieve 67.3% and 41.5% robust accuracy on CIFAR-10 and CIFAR-100 (improving upon the state-of-the-art by +7.23% and +9.07%). We also outperform methods using a training set with a far larger scale than ours. All our models and codes are available online at https://github.com/huangd1999/BORT.
A typical multi-source domain adaptation (MSDA) approach aims to transfer knowledge learned from a set of labeled source domains, to an unlabeled target domain. Nevertheless, prior works strictly assume that each source domain shares the identical group of classes with the target domain, which could hardly be guaranteed as the target label space is not observable. In this paper, we consider a more versatile setting of MSDA, namely Generalized Multi-source Domain Adaptation, wherein the source domains are partially overlapped, and the target domain is allowed to contain novel categories that are not presented in any source domains. This new setting is more elusive than any existing domain adaptation protocols due to the coexistence of the domain and category shifts across the source and target domains. To address this issue, we propose a variational domain disentanglement (VDD) framework, which decomposes the domain representations and semantic features for each instance by encouraging dimension-wise independence. To identify the target samples of unknown classes, we leverage online pseudo labeling, which assigns the pseudo-labels to unlabeled target data based on the confidence scores. Quantitative and qualitative experiments conducted on two benchmark datasets demonstrate the validity of the proposed framework.
Generalized Zero-Shot Learning (GZSL) aims to recognize images from both the seen and unseen classes by transferring semantic knowledge from seen to unseen classes. It is a promising solution to take the advantage of generative models to hallucinate realistic unseen samples based on the knowledge learned from the seen classes. However, due to the generation shifts, the synthesized samples by most existing methods may drift from the real distribution of the unseen data. To address this issue, we propose a novel flow-based generative framework that consists of multiple conditional affine coupling layers for learning unseen data generation. Specifically, we discover and address three potential problems that trigger the generation shifts, i.e., semantic inconsistency, variance collapse, and structure disorder. First, to enhance the reflection of the semantic information in the generated samples, we explicitly embed the semantic information into the transformation in each conditional affine coupling layer. Second, to recover the intrinsic variance of the real unseen features, we introduce a boundary sample mining strategy with entropy maximization to discover more difficult visual variants of semantic prototypes and hereby adjust the decision boundary of the classifiers. Third, a relative positioning strategy is proposed to revise the attribute embeddings, guiding them to fully preserve the inter-class geometric structure and further avoid structure disorder in the semantic space. Extensive experimental results on four GZSL benchmark datasets demonstrate that GSMFlow achieves the state-of-the-art performance on GZSL.
Inspired by the strong ties between vision and language, the two intimate human sensing and communication modalities, our paper aims to explore the generation of 3D human full-body motions from texts, as well as its reciprocal task, shorthanded for text2motion and motion2text, respectively. To tackle the existing challenges, especially to enable the generation of multiple distinct motions from the same text, and to avoid the undesirable production of trivial motionless pose sequences, we propose the use of motion token, a discrete and compact motion representation. This provides one level playing ground when considering both motions and text signals, as the motion and text tokens, respectively. Moreover, our motion2text module is integrated into the inverse alignment process of our text2motion training pipeline, where a significant deviation of synthesized text from the input text would be penalized by a large training loss; empirically this is shown to effectively improve performance. Finally, the mappings in-between the two modalities of motions and texts are facilitated by adapting the neural model for machine translation (NMT) to our context. This autoregressive modeling of the distribution over discrete motion tokens further enables non-deterministic production of pose sequences, of variable lengths, from an input text. Our approach is flexible, could be used for both text2motion and motion2text tasks. Empirical evaluations on two benchmark datasets demonstrate the superior performance of our approach on both tasks over a variety of state-of-the-art methods. Project page: https://ericguo5513.github.io/TM2T/