Scale variation is a deep-rooted problem in object counting, which has not been effectively addressed by existing scale-aware algorithms. An important factor is that they typically involve cooperative learning across multi-resolutions, which could be suboptimal for learning the most discriminative features from each scale. In this paper, we propose a novel method termed STEERER (\textbf{S}elec\textbf{T}iv\textbf{E} inh\textbf{ER}itance l\textbf{E}a\textbf{R}ning) that addresses the issue of scale variations in object counting. STEERER selects the most suitable scale for patch objects to boost feature extraction and only inherits discriminative features from lower to higher resolution progressively. The main insights of STEERER are a dedicated Feature Selection and Inheritance Adaptor (FSIA), which selectively forwards scale-customized features at each scale, and a Masked Selection and Inheritance Loss (MSIL) that helps to achieve high-quality density maps across all scales. Our experimental results on nine datasets with counting and localization tasks demonstrate the unprecedented scale generalization ability of STEERER. Code is available at \url{https://github.com/taohan10200/STEERER}.
In 3D human action recognition, limited supervised data makes it challenging to fully tap into the modeling potential of powerful networks such as transformers. As a result, researchers have been actively investigating effective self-supervised pre-training strategies. In this work, we show that instead of following the prevalent pretext task to perform masked self-component reconstruction in human joints, explicit contextual motion modeling is key to the success of learning effective feature representation for 3D action recognition. Formally, we propose the Masked Motion Prediction (MAMP) framework. To be specific, the proposed MAMP takes as input the masked spatio-temporal skeleton sequence and predicts the corresponding temporal motion of the masked human joints. Considering the high temporal redundancy of the skeleton sequence, in our MAMP, the motion information also acts as an empirical semantic richness prior that guide the masking process, promoting better attention to semantically rich temporal regions. Extensive experiments on NTU-60, NTU-120, and PKU-MMD datasets show that the proposed MAMP pre-training substantially improves the performance of the adopted vanilla transformer, achieving state-of-the-art results without bells and whistles. The source code of our MAMP is available at https://github.com/maoyunyao/MAMP.
Structural re-parameterization is a general training scheme for Convolutional Neural Networks (CNNs), which achieves performance improvement without increasing inference cost. As Vision Transformers (ViTs) are gradually surpassing CNNs in various visual tasks, one may question: if a training scheme specifically for ViTs exists that can also achieve performance improvement without increasing inference cost? Recently, Mixture-of-Experts (MoE) has attracted increasing attention, as it can efficiently scale up the capacity of Transformers at a fixed cost through sparsely activated experts. Considering that MoE can also be viewed as a multi-branch structure, can we utilize MoE to implement a ViT training scheme similar to structural re-parameterization? In this paper, we affirmatively answer these questions, with a new general training strategy for ViTs. Specifically, we decouple the training and inference phases of ViTs. During training, we replace some Feed-Forward Networks (FFNs) of the ViT with specially designed, more efficient MoEs that assign tokens to experts by random uniform partition, and perform Experts Weights Averaging (EWA) on these MoEs at the end of each iteration. After training, we convert each MoE into an FFN by averaging the experts, transforming the model back into original ViT for inference. We further provide a theoretical analysis to show why and how it works. Comprehensive experiments across various 2D and 3D visual tasks, ViT architectures, and datasets validate the effectiveness and generalizability of the proposed training scheme. Besides, our training scheme can also be applied to improve performance when fine-tuning ViTs. Lastly, but equally important, the proposed EWA technique can significantly improve the effectiveness of naive MoE in various 2D visual small datasets and 3D visual tasks.
This paper proposes a novel transformer-based framework that aims to enhance weakly supervised semantic segmentation (WSSS) by generating accurate class-specific object localization maps as pseudo labels. Building upon the observation that the attended regions of the one-class token in the standard vision transformer can contribute to a class-agnostic localization map, we explore the potential of the transformer model to capture class-specific attention for class-discriminative object localization by learning multiple class tokens. We introduce a Multi-Class Token transformer, which incorporates multiple class tokens to enable class-aware interactions with the patch tokens. To achieve this, we devise a class-aware training strategy that establishes a one-to-one correspondence between the output class tokens and the ground-truth class labels. Moreover, a Contrastive-Class-Token (CCT) module is proposed to enhance the learning of discriminative class tokens, enabling the model to better capture the unique characteristics and properties of each class. As a result, class-discriminative object localization maps can be effectively generated by leveraging the class-to-patch attentions associated with different class tokens. To further refine these localization maps, we propose the utilization of patch-level pairwise affinity derived from the patch-to-patch transformer attention. Furthermore, the proposed framework seamlessly complements the Class Activation Mapping (CAM) method, resulting in significantly improved WSSS performance on the PASCAL VOC 2012 and MS COCO 2014 datasets. These results underline the importance of the class token for WSSS.
Model-based reinforcement learning (RL) has demonstrated remarkable successes on a range of continuous control tasks due to its high sample efficiency. To save the computation cost of conducting planning online, recent practices tend to distill optimized action sequences into an RL policy during the training phase. Although the distillation can incorporate both the foresight of planning and the exploration ability of RL policies, the theoretical understanding of these methods is yet unclear. In this paper, we extend the policy improvement step of Soft Actor-Critic (SAC) by developing an approach to distill from model-based planning to the policy. We then demonstrate that such an approach of policy improvement has a theoretical guarantee of monotonic improvement and convergence to the maximum value defined in SAC. We discuss effective design choices and implement our theory as a practical algorithm -- Model-based Planning Distilled to Policy (MPDP) -- that updates the policy jointly over multiple future time steps. Extensive experiments show that MPDP achieves better sample efficiency and asymptotic performance than both model-free and model-based planning algorithms on six continuous control benchmark tasks in MuJoCo.
Multimodal learning aims to build models that can process and relate information from multiple modalities. Despite years of development in this field, it still remains challenging to design a unified network for processing various modalities ($\textit{e.g.}$ natural language, 2D images, 3D point clouds, audio, video, time series, tabular data) due to the inherent gaps among them. In this work, we propose a framework, named Meta-Transformer, that leverages a $\textbf{frozen}$ encoder to perform multimodal perception without any paired multimodal training data. In Meta-Transformer, the raw input data from various modalities are mapped into a shared token space, allowing a subsequent encoder with frozen parameters to extract high-level semantic features of the input data. Composed of three main components: a unified data tokenizer, a modality-shared encoder, and task-specific heads for downstream tasks, Meta-Transformer is the first framework to perform unified learning across 12 modalities with unpaired data. Experiments on different benchmarks reveal that Meta-Transformer can handle a wide range of tasks including fundamental perception (text, image, point cloud, audio, video), practical application (X-Ray, infrared, hyperspectral, and IMU), and data mining (graph, tabular, and time-series). Meta-Transformer indicates a promising future for developing unified multimodal intelligence with transformers. Code will be available at https://github.com/invictus717/MetaTransformer
Temporal modeling plays a crucial role in understanding video content. To tackle this problem, previous studies built complicated temporal relations through time sequence thanks to the development of computationally powerful devices. In this work, we explore the potential of four simple arithmetic operations for temporal modeling. Specifically, we first capture auxiliary temporal cues by computing addition, subtraction, multiplication, and division between pairs of extracted frame features. Then, we extract corresponding features from these cues to benefit the original temporal-irrespective domain. We term such a simple pipeline as an Arithmetic Temporal Module (ATM), which operates on the stem of a visual backbone with a plug-andplay style. We conduct comprehensive ablation studies on the instantiation of ATMs and demonstrate that this module provides powerful temporal modeling capability at a low computational cost. Moreover, the ATM is compatible with both CNNs- and ViTs-based architectures. Our results show that ATM achieves superior performance over several popular video benchmarks. Specifically, on Something-Something V1, V2 and Kinetics-400, we reach top-1 accuracy of 65.6%, 74.6%, and 89.4% respectively. The code is available at https://github.com/whwu95/ATM.
Human intelligence can retrieve any person according to both visual and language descriptions. However, the current computer vision community studies specific person re-identification (ReID) tasks in different scenarios separately, which limits the applications in the real world. This paper strives to resolve this problem by proposing a new instruct-ReID task that requires the model to retrieve images according to the given image or language instructions.Our instruct-ReID is a more general ReID setting, where existing ReID tasks can be viewed as special cases by designing different instructions. We propose a large-scale OmniReID benchmark and an adaptive triplet loss as a baseline method to facilitate research in this new setting. Experimental results show that the baseline model trained on our OmniReID benchmark can improve +0.6%, +1.4%, 0.2% mAP on Market1501, CUHK03, MSMT17 for traditional ReID, +0.8%, +2.0%, +13.4% mAP on PRCC, VC-Clothes, LTCC for clothes-changing ReID, +11.7% mAP on COCAS+ real2 for clothestemplate based clothes-changing ReID when using only RGB images, +25.4% mAP on COCAS+ real2 for our newly defined language-instructed ReID. The dataset, model, and code will be available at https://github.com/hwz-zju/Instruct-ReID.
Generating realistic human motion from given action descriptions has experienced significant advancements because of the emerging requirement of digital humans. While recent works have achieved impressive results in generating motion directly from textual action descriptions, they often support only a single modality of the control signal, which limits their application in the real digital human industry. This paper presents a Motion General-Purpose generaTor (MotionGPT) that can use multimodal control signals, e.g., text and single-frame poses, for generating consecutive human motions by treating multimodal signals as special input tokens in large language models (LLMs). Specifically, we first quantize multimodal control signals into discrete codes and then formulate them in a unified prompt instruction to ask the LLMs to generate the motion answer. Our MotionGPT demonstrates a unified human motion generation model with multimodal control signals by tuning a mere 0.4% of LLM parameters. To the best of our knowledge, MotionGPT is the first method to generate human motion by multimodal control signals, which we hope can shed light on this new direction. Codes shall be released upon acceptance.
The field of generative AI has a transformative impact on various areas, including virtual reality, autonomous driving, the metaverse, gaming, and robotics. Among these applications, 3D object generation techniques are of utmost importance. This technique has unlocked fresh avenues in the realm of creating, customizing, and exploring 3D objects. However, the quality and diversity of existing 3D object generation methods are constrained by the inadequacies of existing 3D object datasets, including issues related to text quality, the incompleteness of multi-modal data representation encompassing 2D rendered images and 3D assets, as well as the size of the dataset. In order to resolve these issues, we present UniG3D, a unified 3D object generation dataset constructed by employing a universal data transformation pipeline on Objaverse and ShapeNet datasets. This pipeline converts each raw 3D model into comprehensive multi-modal data representation <text, image, point cloud, mesh> by employing rendering engines and multi-modal models. These modules ensure the richness of textual information and the comprehensiveness of data representation. Remarkably, the universality of our pipeline refers to its ability to be applied to any 3D dataset, as it only requires raw 3D data. The selection of data sources for our dataset is based on their scale and quality. Subsequently, we assess the effectiveness of our dataset by employing Point-E and SDFusion, two widely recognized methods for object generation, tailored to the prevalent 3D representations of point clouds and signed distance functions. Our dataset is available at: https://unig3d.github.io.