IEEE transactions on pattern analysis and machine intelligence最新文献

Policy diversity, encompassing the variety of policies an agent can adopt, enhances reinforcement learning (RL) success by fostering more robust, adaptable, and innovative problem-solving in the environment. The environment in which standard RL operates is usually modeled with a Markov Decision Process (MDP) as the theoretical foundation. However, in many real-world scenarios, the rewards depend on an agent's history of states and actions leading to a non-MDP. Under the premise of policy diffusion initialization, non-MDPs may have unstructured expanding solution space due to varying historical information and temporal dependencies. This results in solutions having non-equivalent closed forms in non-MDPs. In this paper, deriving diverse solutions for non-MDPs requires policies to break through the boundaries of the current solution space through gradual dispersion. The goal is to expand the solution space, thereby obtaining more diverse policies. Specifically, we first model the sequences of states and actions by a transformer-based method to learn policy embeddings for dispersion in the solution space, since the transformer has advantages in handling sequential data and capturing long-range dependencies for non-MDP. Then, we stack the policy embeddings to construct a dispersion matrix as the policy diversity measure to induce the policy dispersion in the solution space and obtain a set of diverse policies. Finally, we prove that if the dispersion matrix is positive definite, the dispersed embeddings can effectively enlarge the disagreements across policies, yielding a diverse expression for the original policy embedding distribution. Experimental results of both non-MDP and MDP environments show that this dispersion scheme can obtain more expressive diverse policies via expanding the solution space, showing more robust performance than the recent learning baselines.

We propose an end-to-end visuomotor navigation framework that leverages Neural Radiance Fields (NeRF) for spatial cognition. To the best of our knowledge, this is the first effort to integrate such implicit spatial representation with embodied policy end-to-end for cognitive decision-making. Consequently, our system does not necessitate modularized designs nor transformations into explicit scene representations for downstream control. The NeRF-based memory is constructed online during navigation, without relying on any environmental priors. To enhance the extraction of decision-critical historical insights from the rigid and implicit structure of NeRF, we introduce a spatial information extraction mechanism named Structural Radiance Attention (SRA). SRA empowers the agent to grasp complex scene structures and task objectives, thus paving the way for the development of intelligent behavioral patterns. Our comprehensive testing in image-goal navigation tasks demonstrates that our approach significantly outperforms existing navigation models. We demonstrate that SRA markedly improves the agent's understanding of both the scene and the task by retrieving historical information stored in NeRF memory. The agent also learns exploratory awareness from our pipeline to better adapt to low signal-to-noise memory signals in unknown scenes. We deploy our navigation system on a mobile robot in real-world scenarios, where it exhibits evident cognitive capabilities while ensuring real-time performance.

Partial label learning (PLL) is a form of weakly supervised learning, where each training example is linked to a set of candidate labels, among which only one label is correct. Most existing PLL approaches assume that the incorrect labels in each training example are randomly picked as the candidate labels. However, in practice, this assumption may not hold true, as the candidate labels are often instance-dependent. In this paper, we address the instance-dependent PLL problem and assume that each example is associated with a latent label distribution where the incorrect label with a high degree is more likely to be annotated as a candidate label. Motivated by this consideration, we propose two methods VALEN and MILEN, which train the predictive model via utilizing the latent label distributions recovered by the label enhancement process. Specifically, VALEN recovers the latent label distributions via inferring the variational posterior density parameterized by an inference model with the deduced evidence lower bound. MILEN recovers the latent label distribution by adopting the variational approximation to bound the mutual information among the latent label distribution, observed labels and augmented instances. Experiments on benchmark and real-world datasets validate the effectiveness of the proposed methods.

Contrastive Language-Image Pre-training (CLIP) has recently shown great promise in pixel-level zero-shot learning tasks. However, existing approaches utilizing CLIP's text and patch embeddings to generate semantic masks often misidentify input pixels from unseen classes, leading to confusion between novel classes and semantically similar ones. In this work, we propose a novel approach, TagCLIP (Trusty-aware guided CLIP), to address this issue. We disentangle the ill-posed optimization problem into two parallel processes: semantic matching performed individually and reliability judgment for improving discrimination ability. Building on the idea of special tokens in language modeling representing sentence-level embeddings, we introduce a trusty token that enables distinguishing novel classes from known ones in prediction. To evaluate our approach, we conduct experiments on two benchmark datasets, PASCAL VOC 2012 and COCO-Stuff 164K. Our results show that TagCLIP improves the Intersection over Union (IoU) of unseen classes by 7.4% and 1.7%, respectively, with negligible overheads.

In this paper, we introduce Neural Collaborative Search (NCS), a novel learning-based framework for efficiently solving pickup and delivery problems (PDPs). NCS pioneers the collaboration between the latest prevalent neural construction and neural improvement models, establishing a collaborative framework where an improvement model iteratively refines solutions initiated by a construction model. Our NCS collaboratively trains the two models via reinforcement learning with an effective shared-critic mechanism. In addition, the construction model enhances the improvement model with high-quality initial solutions via curriculum learning, while the improvement model accelerates the convergence of the construction model through imitation learning. Besides the new framework design, we also propose the efficient Neural Neighborhood Search (N2S), an efficient improvement model employed within the NCS framework. N2S exploits a tailored Markov decision process formulation and two customized decoders for removing and then reinserting a pair of pickup-delivery nodes, thereby learning a ruin-repair search process for addressing the precedence constraints in PDPs efficiently. To balance the computation cost between encoders and decoders, N2S streamlines the existing encoder design through a light Synthesis Attention mechanism that allows the vanilla self-attention to synthesize various features regarding a route solution. Moreover, a diversity enhancement scheme is further leveraged to ameliorate the performance during the inference of N2S. Our NCS and N2S are both generic, and extensive experiments on two canonical PDP variants show that they can produce state-of-the-art results among existing neural methods. Remarkably, our NCS and N2S could surpass the well-known LKH3 solver especially on the more constrained PDP variant.

Panoptic Part Segmentation (PPS) unifies panoptic and part segmentation into one task. Previous works utilize separate approaches to handle things, stuff, and part predictions without shared computation and task association. We aim to unify these tasks at the architectural level, designing the first end-to-end unified framework, Panoptic-PartFormer. Moreover, we find the previous metric PartPQ biases to PQ. To handle both issues, we first design a meta-architecture that decouples part features and things/stuff features, respectively. We model things, stuff, and parts as object queries and directly learn to optimize all three forms of prediction as a unified mask prediction and classification problem. We term our model as Panoptic-PartFormer. Second, we propose a new metric Part-Whole Quality (PWQ), better to measure this task from pixel-region and part-whole perspectives. It also decouples the errors for part segmentation and panoptic segmentation. Third, inspired by Mask2Former, based on our meta-architecture, we propose Panoptic-PartFormer++ and design a new part-whole cross-attention scheme to boost part segmentation qualities further. We design a new part-whole interaction method using masked cross attention. Finally, extensive ablation studies and analysis demonstrate the effectiveness of both Panoptic-PartFormer and Panoptic-PartFormer++. Compared with previous Panoptic-PartFormer, our Panoptic-PartFormer++ achieves 2% PartPQ and 3% PWQ improvements on the Cityscapes PPS dataset and 5% PartPQ on the Pascal Context PPS dataset. On both datasets, Panoptic-PartFormer++ achieves new state-of-the-art results. Our models can serve as a strong baseline and aid future research in PPS. The source code and trained models will be available at https://github.com/lxtGH/Panoptic-PartFormer.

Topological data analysis provides a set of tools to uncover low-dimensional structure in noisy point clouds. Prominent amongst the tools is persistence homology, which summarizes birth-death times of homological features using data objects known as persistence diagrams. To better aid statistical analysis, a functional representation of the diagrams, known as persistence landscapes, enable use of functional data analysis and machine learning tools. Topological and geometric variabilities inherent in point clouds are confounded in both persistence diagrams and landscapes, and it is important to distinguish topological signal from noise to draw reliable conclusions on the structure of the point clouds when using persistence homology. We develop a framework for decomposing variability in persistence diagrams into topological signal and topological noise through alignment of persistence landscapes using an elastic Riemannian metric. Aligned landscapes (amplitude) isolate the topological signal. Reparameterizations used for landscape alignment (phase) are linked to a resolution parameter used to generate persistence diagrams, and capture topological noise in the form of geometric, global scaling and sampling variabilities. We illustrate the importance of decoupling topological signal and topological noise in persistence diagrams (landscapes) using several simulated examples. We also demonstrate that our approach provides novel insights in two real data studies.

Image- and video-based 3D human recovery (i.e., pose and shape estimation) have achieved substantial progress. However, due to the prohibitive cost of motion capture, existing datasets are often limited in scale and diversity. In this work, we obtain massive human sequences by playing the video game with automatically annotated 3D ground truths. Specifically, we contribute GTA-Human, a large-scale 3D human dataset generated with the GTA-V game engine, featuring a highly diverse set of subjects, actions, and scenarios. More importantly, we study the use of game-playing data and obtain five major insights. First, game-playing data is surprisingly effective. A simple frame-based baseline trained on GTA-Human outperforms more sophisticated methods by a large margin. For videobased methods, GTA-Human is even on par with the in-domain training set. Second, we discover that synthetic data provides critical complements to the real data that is typically collected indoor. We highlight that our investigation into domain gap provides explanations for our data mixture strategies that are simple yet useful, which offers new insights to the research community. Third, the scale of the dataset matters. The performance boost is closely related to the additional data available. A systematic study on multiple key factors (such as camera angle and body pose) reveals that the model performance is sensitive to data density. Fourth, the effectiveness of GTA-Human is also attributed to the rich collection of strong supervision labels (SMPL parameters), which are otherwise expensive to acquire in real datasets. Fifth, the benefits of synthetic data extend to larger models such as deeper convolutional neural networks (CNNs) and Transformers, for which a significant impact is also observed. We hope our work could pave the way for scaling up 3D human recovery to the real world. Homepage: https://caizhongang.github.io/projects/GTA-Human/.

Video-to-Video synthesis (Vid2Vid) gains remarkable performance in generating a photo-realistic video from a sequence of semantic maps, such as segmentation, sketch and pose. However, this pipeline is heavily limited to high computational cost and long inference latency, mainly attributed to two essential factors: 1) network architecture parameters, 2) sequential data stream. Recently, the parameters of image-based generative models have been significantly reduced via more efficient network architectures. Existing methods mainly focus on slimming network architectures but ignore the size of the sequential data stream. Moreover, due to the lack of temporal coherence, image-based compression is not sufficient for the compression of the video task. In this paper, we present a spatial-temporal hybrid distillation compression framework, Fast-Vid2Vid++, which focuses on knowledge distillation of the teacher network and the data stream of generative models on both space and time. Fast-Vid2Vid++ makes the first attempt at time dimension to transfer hierarchical features and time coherence knowledge to reduce computational resources and accelerate inference. Specifically, we compress the data stream spatially and reduce the temporal redundancy. We distill the knowledge of the hierarchical features and the final response from the teacher network to the student network in high-resolution and full-time domains. We transfer the long-term dependencies of the features and video frames to the student model. After the proposed spatial-temporal hybrid knowledge distillation (Spatial-Temporal-HKD), our model can synthesize high-resolution key-frames using the low-resolution data stream. Finally, Fast-Vid2Vid++ interpolates intermediate frames by motion compensation with slight latency and generates full-length sequences with motion-aware inference (MAI). On standard benchmarks, Fast-Vid2Vid++ achieves a real-time performance of 30-59 FPS and saves 28-35× computational cost on a single V100 GPU. Code and models are publicly available.

DeepTensor is a computationally efficient framework for low-rank decomposition of matrices and tensors using deep generative networks. We decompose a tensor as the product of low-rank tensor factors (e.g., a matrix as the outer product of two vectors), where each low-rank tensor is generated by a deep network (DN) that is trained in a self-supervised manner to minimize the mean-square approximation error. Our key observation is that the implicit regularization inherent in DNs enables them to capture nonlinear signal structures (e.g., manifolds) that are out of the reach of classical linear methods like the singular value decomposition (SVD) and principal components analysis (PCA). Furthermore, in contrast to the SVD and PCA, whose performance deteriorates when the tensor's entries deviate from additive white Gaussian noise, we demonstrate that the performance of DeepTensor is robust to a wide range of distributions. We validate that DeepTensor is a robust and computationally efficient drop-in replacement for the SVD, PCA, nonnegative matrix factorization (NMF), and similar decompositions by exploring a range of real-world applications, including hyperspectral image denoising, 3D MRI tomography, and image classification. In particular, DeepTensor offers a 6dB signal-to-noise ratio improvement over standard denoising methods for signal corrupted by Poisson noise and learns to decompose 3D tensors 60 times faster than a single DN equipped with 3D convolutions.