Workflow
电子游戏
icon
Search documents
ACMMM 2025 | 北大团队提出 InteractMove:3D场景中人与可移动物体交互动作生成新框架
机器之心· 2025-10-19 03:48
Core Insights - The article introduces the research paper titled "InteractMove: Text-Controlled Human-Object Interaction Generation in 3D Scenes with Movable Objects," which presents a novel task of generating human-object interactions in 3D scenes based on text descriptions, specifically focusing on movable objects [3][7][35] - The research team has developed a large-scale dataset and an innovative framework that outperforms existing methods in various evaluation metrics, addressing the limitations of current human-scene interaction datasets that primarily focus on static objects [3][4][35] Dataset Highlights - The InteractMove dataset includes multiple interactive objects and interference items, requiring the model to understand language and spatial reasoning to select the correct object [11] - It covers 71 types of movable objects and 21 interaction methods, ensuring a diverse range of interactions from simple to complex [11] - The dataset ensures physical realism by rigorously filtering actions and trajectories to avoid unrealistic phenomena like "penetration" [11][12] Methodology Overview - The proposed framework consists of three core modules: 3D visual localization, hand-object reachability graph learning, and collision-aware action generation [20][21][22] - The first step involves accurately locating target objects in complex scenes based on text input [20] - The second step models the fine-grained contact relationships between hand joints and object surfaces, allowing for diverse interaction strategies [21] - The final step ensures that generated actions adhere to physical laws, preventing collisions and ensuring natural interactions [22][23] Experimental Results - The method demonstrates superior performance across all key metrics, including interaction accuracy, physical realism, diversity, and collision avoidance, with a 18% improvement in diversity and a 14% improvement in physical realism compared to the best existing results [24][25] - Ablation studies confirm the effectiveness and necessity of each module in the proposed framework [28][29] Qualitative Analysis - The visual results indicate that InteractMove generates semantically coherent, natural, and physically realistic human-object interactions, showcasing smooth action transitions and appropriate hand-object contact [31][32][33] - The generated actions align closely with human-like behavior, avoiding unrealistic poses and ensuring that object movements are coordinated with human actions [32][33] Conclusion - The InteractMove project establishes a new framework for text-driven human-object interaction generation, overcoming the limitations of static object interactions and laying a solid foundation for applications in virtual reality, augmented reality, digital humans, and robotics [35]
早报|大疆就诉美国防部判决结果提起上诉;80名韩国公民被柬拘留拒绝回国;2026年国考报考年龄放宽;京东回应下场造车
虎嗅APP· 2025-10-15 00:01
Group 1 - 80 South Korean citizens are currently detained by the Cambodian immigration authorities, and they refuse to return to South Korea despite contact from officials [2][3][5] - An semiconductor company, Anshi Semiconductor, appointed its CFO Stefan Tilger as interim CEO after the previous CEO was suspended by a Dutch court [6] - Goldman Sachs plans to conduct another round of layoffs to further cut costs, although the total number of employees is expected to grow by the end of the year [7] Group 2 - Walmart announced a partnership with OpenAI to allow customers to use ChatGPT for instant checkout of Walmart products [8] - The Chinese Embassy in Thailand issued a warning regarding a tourist being threatened to shop, confirming that the involved guide lacked proper qualifications [9][10][11] - DJI has appealed a US court decision that maintained its listing as a "Chinese military enterprise," asserting its commitment to preventing the misuse of its products for military purposes [12] Group 3 - Ele.me is trialing a new service score system to replace the previous penalty system for delivery riders, aiming for a more positive incentive approach [13] - The 2026 national civil service examination will recruit 38,100 candidates, with age limits adjusted for applicants [14] - The total number of vehicle trade-ins is expected to exceed 12 million this year, significantly boosting new car sales [15][17] Group 4 - Nvidia's CEO personally delivered the first DGX Spark AI supercomputers to Elon Musk, highlighting advancements in AI computing capabilities [18] - JD.com is collaborating with CATL and GAC Group to launch a new car, clarifying that JD's role is primarily in consumer insights and sales, not manufacturing [21] - Apple CEO Tim Cook's visit to China coincided with reports of activation issues with the iPhone series, attributed to server problems [23] Group 5 - Ant Group reported a surge in interest for its gold accumulation product, with over a million visits on the day international gold prices hit a record high [24] - Songcheng Performance announced it is not involved in the IPO project of Xibei and has no plans to enter the prepared food industry [25][28] - The photovoltaic industry association clarified that its upcoming meetings are routine monthly meetings, amidst speculation of significant policy announcements [29][30] Group 6 - The central bank is set to conduct a 600 billion yuan reverse repurchase operation to maintain liquidity in the banking system [34] - Elon Musk indicated that his company xAI will focus on video games, driven by personal interest rather than profit [35]