Home Technology peripherals AI CVPR 2024 | Is there only single-person data in the synthetic video data set? M3Act solves the problem of crowd behavior labeling

CVPR 2024 | Is there only single-person data in the synthetic video data set? M3Act solves the problem of crowd behavior labeling

Jun 03, 2024 pm 10:02 PM
industry M3Act Synthetic data generation framework

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题
The AIxiv column is a column where this site publishes academic and technical content. In the past few years, the AIxiv column of this site has received more than 2,000 reports, covering top laboratories from major universities and companies around the world, effectively promoting academic exchanges and dissemination. If you have excellent work that you want to share, please feel free to contribute or contact us for reporting. Submission email: liyazhou@jiqizhixin.com; zhaoyunfeng@jiqizhixin.com

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

  • ## Paper link: https://arxiv. org/abs/2306.16772
  • Project link: https://cjerry1243.github.io/M3Act/
  • Paper title: M3Act: Learning from Synthetic Human Group Activities

Introduction

Recognizing and understanding crowd behavior through visual information is an important area in video monitoring, interactive robots, autonomous driving and other fields is one of the key technologies, but obtaining large-scale crowd behavior annotation data has become a bottleneck in the development of related research. Nowadays, synthetic datasets are becoming an emerging method to replace real-world data, but synthetic datasets in existing research mainly focus on the estimation of human pose and shape. They often only provide synthetic animation videos of
single characters, which are not suitable for video recognition tasks of crowds.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

In this article, the author proposes M3Act, a synthetic data generation framework suitable for multi-group crowd behavior. Experiments show that this synthetic data set can greatly improve the performance of downstream models in multi-person tracking and group activity recognition, and can replace more than 62.5% of real data on the DanceTrack task, thereby reducing data annotation costs in real-world application scenarios. Additionally, this synthetic data framework proposes a new class of tasks: controllable 3D swarm activity generation. This task aims to directly control the swarm activity generation results using multiple inputs (activity category, swarm size, trajectory, density, speed, and text input). The authors rigorously define tasks and metrics and provide competitive baselines and results.

Data generation
Based on the Unity engine development, M3Act covers a variety of behavioral types of crowd data , provides highly diverse and realistic video images, as well as comprehensive data labeling. Compared to other synthetic datasets, M3Act provides more comprehensive labeled data, including 2D and 3D markers as well as fine-grained individual-level and group-level labels, thus making it an ideal synthesis to support multi-person and multi-group research tasks Dataset generator.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

Data generator includes 25 3D scenes, 104 high dynamic range panoramic images, 5 light settings, 2200 character models, 384 animations (14 action categories ) and 6 group activity types. The data generation process is as follows. First, all parameters within a simulation scenario are determined through a randomization process, and then a 3D scene with background objects, lights and cameras, and a group of character models with animation are generated based on the parameters. Finally, RGB images are rendered from multiple viewpoints and the labeled results are exported.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

To ensure a high degree of diversity in simulated data, M3Act provides randomization for nearly all aspects of the data generation process. This includes the number of groups in the scene, the number of people in each group, the position of the group, the arrangement of people in the group, the position of the individuals, the textures of the instantiated characters, as well as the scene, lighting conditions, camera position, characters, group activity, atoms Selection of action and animation clips. Each group activity is also built as a parameterized module. These parameters include the number of individuals in the swarm and the specific atomic actions allowed within the swarm's activity.

The final generated data set is divided into two parts. The first part "M3ActRGB" contains 6000 simulations of single but multiple types of group activities and 9000 simulations of multiple groups and multiple types, with a total of 6 million RGB images and 48 million bounding boxes. The second part "M3Act3D" contains only 3D data. It consists of more than 65,000 150-frame simulations of a single multi-type group activity, totaling 87.6 hours. To the authors' knowledge, M3Act3D's group size and interaction complexity are significantly higher than previous multiplayer sports datasets, making it the first large-scale 3D dataset for large group activities.

Experimental results

The actual effect of M3Act is through three The core experiments demonstrate: multi-person tracking, group activity recognition and controllable group activity generation.

Experiment 1: Multi-person Tracking

Research findings , after adding synthetic data to the training of the existing model MOTRv2 [1], the model has significantly improved on all 5 indicators, especially jumping from 10th to 2nd in the ranking on the HOTA indicator. At the same time, when 62.5% of the real data in the training set was replaced by synthetic data, the model could still achieve similar performance. In addition, compared to other synthetic data sources, such as BEDLAM and GTA-Humans, M3Act provides greater performance improvements for model training, indicating that it is more suitable for multi-person group activity tasks. Finally, the table below shows the training results of different models under M3Act. The results show that M3Act is effective in various models.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

Experiment 2: Group activity recognition

Similarly , M3Act also improves the performance of two existing group activity recognition models, as shown in the following table: As the amount of synthetic data used for pre-training increases, the recognition accuracy continues to improve. When using 100% synthetic data, the accuracy of the group activity recognition model Composer [2] increased by an average of 4.87% at the group level and 7.43% at the individual level, while another group activity recognition model Actor Transformer [3] improved at the group level. An increase of 5.59% in accuracy was seen on , and an increase of 5.43% at the individual level.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

The following table shows the group recognition accuracy on CAD2 and Volleyball (VD) using different input modalities. Performance gains in experiments demonstrate that M3Act's synthetic data can effectively benefit downstream tasks and span different models, input modalities, and datasets.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

Experiment 3: Controllable 3D group activity generation

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

The author proposes a new type of task: controllable 3D group activity generation. The task aims to synthesize a set of 3D human actions from Gaussian noise based on a given activity class label and an arbitrary population size. Although existing studies can generate multi-player actions, they are limited to two-person scenarios or groups with a fixed number of people. Therefore, the authors propose two baseline methods. In the first baseline approach, group activity is implemented by repeatedly invoking the single-person motion diffusion model MDM [4], so the generation process for each individual is independent. The second method adds an interactive transformer (IFormer) based on MDM. Due to its modeling of human interactions, MDM+IFormer is able to produce coordinated group activities in a single forward pass.

The author considers the following evaluation indicators at both the group and individual levels: recognition accuracy, Frechette initial distance (FID), diversity and multimodality. In addition, based on the social force model, the author adds four location-based indicators at the group level: collision frequency, repulsive interaction force, contact repulsive force, and total repulsive force. The results show:

  • MDM+IFormer is able to generate group activities with well-aligned character positions. See qualitative graph below.
  • Both baseline methods can generate diverse activities matching the input conditions, but MDM+IFormer achieves better FID scores.
  • Interaction transformers in MDM+IFormer greatly reduce the frequency of collisions within generated group activities.

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

CVPR 2024 | 合成视频数据集里只有单人数据?M3Act破解人群行为标注难题

in conclusion

The authors of the paper demonstrate the advantages of M3Act through three core experiments of multi-modality and enhanced performance, as well as the introduction of a new generation task. In experiments on multi-person tracking and group activity recognition, they observed that the model's generalization ability to unseen test cases improved as more synthetic data was added.

In addition, the synthetic data in M3Act can replace real data in some target fields without affecting performance, which is expected to reduce the need for a large amount of real data during the training process, thereby reducing cost of data collection and annotation. This finding demonstrates the potential of small or even zero samples to migrate from simulated data to real-world data.

In the generation of controllable 3D group activities, although MDM+IFormer is only the baseline model for this task, it still learns the interaction rules of character movement and under control Generate well-aligned group activity. Notably, although generative approaches currently outperform procedural approaches, they demonstrate the potential to control group actions directly from a variety of signals (activity category, group size, trajectory, density, speed, and text input). As data availability increases and generative model capabilities improve in the future, the authors predict that generative methods will eventually gain dominance and become more widely used in social interactions and collective human activities.

Although the complexity of group behavior in the M3Act dataset may be limited by the heuristic rules used in the data generation process, M3Act provides significant flexibility in incorporating new group activities , thereby adapting to any specific downstream task. These new populations can originate from expert-guided heuristic rules, rules generated by large language models, or the output of controllable 3D population activity generation models. Furthermore, the paper's authors recognize the domain differences that exist between synthetic and real-world data. With the addition of assets in the data generator in future releases, it will be possible to improve the model's generalization capabilities and mitigate these differences.

[1] Yuang Zhang, Tiancai Wang, and Xiangyu Zhang. Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22056–22065, 2023.
##[2] Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, and Hans Peter Graf. Composer: Compositional reasoning of group activity in videos with keypoint-only modality. Proceedings of the 17th European Conference on Computer Vision (ECCV 2022 ), 2022.
[3] Kirill Gavrilyuk, Ryan Sanford, Mehrsan Javan, and Cees GM Snoek. Actor-transformers for group activity recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 839–848, 2020.
[4] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.

The above is the detailed content of CVPR 2024 | Is there only single-person data in the synthetic video data set? M3Act solves the problem of crowd behavior labeling. For more information, please follow other related articles on the PHP Chinese website!

Statement of this Website
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

Video Face Swap

Video Face Swap

Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Tools

Notepad++7.3.1

Notepad++7.3.1

Easy-to-use and free code editor

SublimeText3 Chinese version

SublimeText3 Chinese version

Chinese version, very easy to use

Zend Studio 13.0.1

Zend Studio 13.0.1

Powerful PHP integrated development environment

Dreamweaver CS6

Dreamweaver CS6

Visual web development tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners Aug 09, 2024 pm 04:01 PM

But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home Aug 21, 2024 pm 07:33 PM

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

Claude has become lazy too! Netizen: Learn to give yourself a holiday Claude has become lazy too! Netizen: Learn to give yourself a holiday Sep 02, 2024 pm 01:56 PM

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded Aug 22, 2024 pm 10:35 PM

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI ​​robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI ​​side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award Aug 15, 2024 pm 04:37 PM

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Aug 08, 2024 am 07:02 AM

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Aug 22, 2024 pm 08:02 PM

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au

Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Sep 03, 2024 pm 05:18 PM

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

See all articles