Multimodal large language models (MLLMs) have enabled open-world visual understanding by injecting visual input as extra tokens into large language models (LLMs) as contexts. However, when the visual input changes from a single image to a long video, the above paradigm encounters difficulty because the vast amount of video tokens has significantly exceeded the maximal capacity of MLLMs. Therefore, existing video-based MLLMs are mostly established upon sampling a small portion of tokens from input data, which can cause key information to be lost and thus produce incorrect answers. This paper presents a simple yet effective algorithm named Adaptive Keyframe Sampling (AKS). It inserts a plug-and-play module known as keyframe selection, which aims to maximize the useful information with a fixed number of video tokens. We formulate keyframe selection as an optimization involving (1) the relevance between the keyframes and the prompt, and (2) the coverage of the keyframes over the video, and present an adaptive algorithm to approximate the best solution. Experiments on two long video understanding benchmarks validate that Adaptive Keyframe Sampling improves video QA accuracy (beyond strong baselines) upon selecting informative keyframes. Our study reveals the importance of information pre-filtering in video-based MLLMs. Code is available at https://github.com/ncTimTang/AKS.

3D human pose estimation has wide applications in fields such as intelligent surveillance, motion capture, and virtual reality. However, in real-world scenarios, issues such as occlusion, noise interference, and missing viewpoints can severely affect pose estimation. To address these challenges, we introduce the task of Deficiency-Aware 3D Pose Estimation. Traditional 3D pose estimation methods often rely on multi-stage networks and modular combinations, which can lead to cumulative errors and increased training complexity, making them unable to effectively address deficiency-aware estimation. To this end, we propose DeProPose, a flexible method that simplifies the network architecture to reduce training complexity and avoid information loss in multi-stage designs. Additionally, the model innovatively introduces a multi-view feature fusion mechanism based on relative projection error, which effectively utilizes information from multiple viewpoints and dynamically assigns weights, enabling efficient integration and enhanced robustness to overcome deficiency-aware 3D Pose Estimation challenges. Furthermore, to thoroughly evaluate this end-to-end multi-view 3D human pose estimation model and to advance research on occlusion-related challenges, we have developed a novel 3D human pose estimation dataset, termed the Deficiency-Aware 3D Pose Estimation (DA-3DPE) dataset. This dataset encompasses a wide range of deficiency scenarios, including noise interference, missing viewpoints, and occlusion challenges. Compared to state-of-the-art methods, DeProPose not only excels in addressing the deficiency-aware problem but also shows improvement in conventional scenarios, providing a powerful and user-friendly solution for 3D human pose estimation. The source code will be available at https://github.com/WUJINHUAN/DeProPose.

Large Language Models (LLMs) have shown impressive reasoning capabilities in well-defined problems with clear solutions, such as mathematics and coding. However, they still struggle with complex real-world scenarios like business negotiations, which require strategic reasoning-an ability to navigate dynamic environments and align long-term goals amidst uncertainty. Existing methods for strategic reasoning face challenges in adaptability, scalability, and transferring strategies to new contexts. To address these issues, we propose explicit policy optimization (EPO) for strategic reasoning, featuring an LLM that provides strategies in open-ended action space and can be plugged into arbitrary LLM agents to motivate goal-directed behavior. To improve adaptability and policy transferability, we train the strategic reasoning model via multi-turn reinforcement learning (RL) using process rewards and iterative self-play, without supervised fine-tuning (SFT) as a preliminary step. Experiments across social and physical domains demonstrate EPO's ability of long-term goal alignment through enhanced strategic reasoning, achieving state-of-the-art performance on social dialogue and web navigation tasks. Our findings reveal various collaborative reasoning mechanisms emergent in EPO and its effectiveness in generating novel strategies, underscoring its potential for strategic reasoning in real-world applications.

Decision making demands intricate interplay between perception, memory, and reasoning to discern optimal policies. Conventional approaches to decision making face challenges related to low sample efficiency and poor generalization. In contrast, foundation models in language and vision have showcased rapid adaptation to diverse new tasks. Therefore, we advocate for the construction of foundation agents as a transformative shift in the learning paradigm of agents. This proposal is underpinned by the formulation of foundation agents with their fundamental characteristics and challenges motivated by the success of large language models (LLMs). Moreover, we specify the roadmap of foundation agents from large interactive data collection or generation, to self-supervised pretraining and adaptation, and knowledge and value alignment with LLMs. Lastly, we pinpoint critical research questions derived from the formulation and delineate trends for foundation agents supported by real-world use cases, addressing both technical and theoretical aspects to propel the field towards a more comprehensive and impactful future.

3D Gaussian splatting has demonstrated impressive performance in real-time novel view synthesis. However, achieving successful reconstruction from RGB images generally requires multiple input views captured under static conditions. To address the challenge of sparse input views, previous approaches have incorporated depth supervision into the training of 3D Gaussians to mitigate overfitting, using dense predictions from pretrained depth networks as pseudo-ground truth. Nevertheless, depth predictions from monocular depth estimation models inherently exhibit significant uncertainty in specific areas. Relying solely on pixel-wise L2 loss may inadvertently incorporate detrimental noise from these uncertain areas. In this work, we introduce a novel method to supervise the depth distribution of 3D Gaussians, utilizing depth priors with integrated uncertainty estimates. To address these localized errors in depth predictions, we integrate a patch-wise optimal transport strategy to complement traditional L2 loss in depth supervision. Extensive experiments conducted on the LLFF, DTU, and Blender datasets demonstrate that our approach, UGOT, achieves superior novel view synthesis and consistently outperforms state-of-the-art methods.

3D human pose estimation captures the human joint points in three-dimensional space while keeping the depth information and physical structure. That is essential for applications that require precise pose information, such as human-computer interaction, scene understanding, and rehabilitation training. Due to the challenges in data collection, mainstream datasets of 3D human pose estimation are primarily composed of multi-view video data collected in laboratory environments, which contains rich spatial-temporal correlation information besides the image frame content. Given the remarkable self-attention mechanism of transformers, capable of capturing the spatial-temporal correlation from multi-view video datasets, we propose a multi-stage framework for 3D sequence-to-sequence (seq2seq) human pose detection. Firstly, the spatial module represents the human pose feature by intra-image content, while the frame-image relation module extracts temporal relationships and 3D spatial positional relationship features between the multi-perspective images. Secondly, the self-attention mechanism is adopted to eliminate the interference from non-human body parts and reduce computing resources. Our method is evaluated on Human3.6M, a popular 3D human pose detection dataset. Experimental results demonstrate that our approach achieves state-of-the-art performance on this dataset.

Decision-making is a dynamic process requiring Self-supervised pretraining has enabled large sequence perception, memory, and reasoning to make models to realize few-shot or even zero-shot adaptation in choices and find optimal policies. Traditional natural language processing (NLP) [OpenAI, 2023] and computer approaches to decision-making suffer from sample vision (CV) tasks [Bai et al., 2023]. Through pretraining efficiency and generalization, while largescale on large generic corpora or visual data (images and self-supervised pretraining has enabled fast videos), knowledge about the world and human society is adaptation with fine-tuning or few-shot learning learned which can be utilized in various downstream task in language and vision. We thus argue to integrate learning with few samples so as to improve sample efficiency knowledge acquired from generic largescale and generalization.

We propose integrally pre-trained transformer pyramid network (iTPN), towards jointly optimizing the network backbone and the neck, so that transfer gap between representation models and downstream tasks is minimal. iTPN is born with two elaborated designs: 1) The first pre-trained feature pyramid upon vision transformer (ViT). 2) Multi-stage supervision to the feature pyramid using masked feature modeling (MFM). iTPN is updated to Fast-iTPN, reducing computational memory overhead and accelerating inference through two flexible designs. 1) Token migration: dropping redundant tokens of the backbone while replenishing them in the feature pyramid without attention operations. 2) Token gathering: reducing computation cost caused by global attention by introducing few gathering tokens. The base/large-level Fast-iTPN achieve 88.75%/89.5% top-1 accuracy on ImageNet-1K. With 1x training schedule using DINO, the base/large-level Fast-iTPN achieves 58.4%/58.8% box AP on COCO object detection, and a 57.5%/58.7% mIoU on ADE20K semantic segmentation using MaskDINO. Fast-iTPN can accelerate the inference procedure by up to 70%, with negligible performance loss, demonstrating the potential to be a powerful backbone for downstream vision tasks. The code is available at: github.com/sunsmarterjie/iTPN.

Backdoor attacks in reinforcement learning (RL) have previously employed intense attack strategies to ensure attack success. However, these methods suffer from high attack costs and increased detectability. In this work, we propose a novel approach, BadRL, which focuses on conducting highly sparse backdoor poisoning efforts during training and testing while maintaining successful attacks. Our algorithm, BadRL, strategically chooses state observations with high attack values to inject triggers during training and testing, thereby reducing the chances of detection. In contrast to the previous methods that utilize sample-agnostic trigger patterns, BadRL dynamically generates distinct trigger patterns based on targeted state observations, thereby enhancing its effectiveness. Theoretical analysis shows that the targeted backdoor attack is always viable and remains stealthy under specific assumptions. Empirical results on various classic RL tasks illustrate that BadRL can substantially degrade the performance of a victim agent with minimal poisoning efforts 0.003% of total training steps) during training and infrequent attacks during testing.

Conventional networks for object skeleton detection are usually hand-crafted. Although effective, they require intensive priori knowledge to configure representative features for objects in different scale granularity.In this paper, we propose adaptive linear span network (AdaLSN), driven by neural architecture search (NAS), to automatically configure and integrate scale-aware features for object skeleton detection. AdaLSN is formulated with the theory of linear span, which provides one of the earliest explanations for multi-scale deep feature fusion. AdaLSN is materialized by defining a mixed unit-pyramid search space, which goes beyond many existing search spaces using unit-level or pyramid-level features.Within the mixed space, we apply genetic architecture search to jointly optimize unit-level operations and pyramid-level connections for adaptive feature space expansion. AdaLSN substantiates its versatility by achieving significantly higher accuracy and latency trade-off compared with state-of-the-arts. It also demonstrates general applicability to image-to-mask tasks such as edge detection and road extraction. Code is available at \href{https://github.com/sunsmarterjie/SDL-Skeleton}{\color{magenta}github.com/sunsmarterjie/SDL-Skeleton}.