摘要

本文介绍了一个名为SoccerKDNet的知识蒸馏框架，用于足球视频中的动作识别。研究者针对传统复杂离线网络难以在资源受限场景部署的问题，提出了一种基于知识蒸馏的端到端迁移学习网络，该网络在Kinetics400数据集上预训练，并引入了独特的损失参数化方法进行分析。文章还介绍了一个新数据集SoccerDB1，包含448个视频，分为4个不同类别的足球运动员动作。研究者提出了一个独特的损失参数，可以线性加权每个网络预测的使用程度，并通过调整各种超参数进行了全面的性能研究。在SoccerDB1数据集上，该模型获得了67.20%的验证准确率，不仅显著优于现有方法，还能轻松泛化到新数据集。该数据集已在https://bit.ly/soccerdb1公开可用。

与SoccerDB1数据集相关的信息

根据网页内容，SoccerDB1是研究者新引入的足球视频数据集，具有以下特点：

1. 数据规模：包含448个视频
2. 类别结构：由4个不同类别的足球运动员动作组成
3. 用途：用于足球视频中的动作识别研究
4. 基准性能：使用SoccerKDNet模型在该数据集上获得了67.20%的验证准确率，这是该数据集上的首个分类基准结果
5. 公开可用性：数据集已在https://bit.ly/soccerdb1公开发布

关于数据集的构建方法，网页内容没有提供详细信息，仅提到它是由研究者新引入的，包含了4个不同类别的足球运动员动作视频。网页没有具体说明数据收集方式、预处理步骤、标注方法等构建细节。

相关网页链接

1. v1 - 论文的第一个版本链接
2. Sarosij Bose - 第一作者链接
3. Saikat Sarkar - 第二作者链接
4. Amlan Chakrabarti - 第三作者链接
5. this https URL - SoccerDB1数据集的下载链接
6. arXiv:2307.07768 - 论文引用链接
7. arXiv:2307.07768v2 - 论文当前版本引用链接
8. https://doi.org/10.48550/arXiv.2307.07768 - 论文DOI链接

相关图片

无与问题相关的图片内容。

SoccerKDNet: A Knowledge Distillation

Framework for Action Recognition in Soccer

Videos

Sarosij Bose1[0000*−0003−3014−4796], Saikat Sarkar2[0000−0001−7118−*6058], and

Amlan Chakrabarti3[0000*−0003−4380−*3172]

1 Department of Computer Science and Engineering, University of Calcutta, India

2 Department of Computer Science, Bangabasi College, University of Calcutta, India

3 A.K.Choudhury School of Information Technology, University of Calcutta, India

{sarosijbose2000, to.saikatsarkar17}@gmail.com,acakcs@caluniv.ac.in

Abstract. Classifying player actions from soccer videos is a challenging

problem, which has become increasingly important in sports analytics

over the years. Most state-of-the-art methods employ highly complex of-

fline networks, which makes it difficult to deploy such models in resource

constrained scenarios. Here, in this paper we propose a novel end-to-end

knowledge distillation based transfer learning network pre-trained on the

Kinetics400 dataset and then perform extensive analysis on the learned

framework by introducing a unique loss parameterization. We also in-

troduce a new dataset named "SoccerDB1" containing 448 videos and

consisting of 4 diverse classes each of players playing soccer. Further-

more, we introduce an unique loss parameter that help us linearly weigh

the extent to which the predictions of each network are utilized. Finally,

we also perform a thorough performance study using various changed

hyperparameters. We also benchmark the first classification results on

the new SoccerDB1 dataset obtaining 67.20% validation accuracy. The

dataset has been made publicly available at:https://bit.ly/soccerdb1

Keywords: Soccer Analytics · Knowledge Distillation · Action Recog-

nition.

1

Introduction

Recognition of player actions in soccer video is a challenging computer vision

task. Existing vision-based soccer analytics models either rely heavily on man-

power which is responsible for tracking every aspect of the game or other offline

network based analytics products that are used for analyzing the game closely

once it’s over [12], [11]. It has been established that deep learning based methods

exceed their traditional counterparts in performance. Recently, deep reinforce-

ment learning based models [14], [13] have been used for the estimation of ball

possession statistics in broadcast soccer videos. But, there is an issue regarding

employing such deep networks, which are often trained on large image based

datasets such as ImageNet. These offline models may deliver superior accuracy

arXiv:2307.07768v2 [cs.CV] 22 Jul 2023

2

Sarosij Bose et al.

but suffers due to a significant domain gap. As a result, there is a need for

domain specific data or, at the very least, fine tuning on sports specific datasets.

Soccer Action Recognition: We also look into the existing literature in

the action recognition domain for soccer videos. However, work has been very

limited regarding this aspect. One of the very few public soccer video datasets is

the Soccernet v2 benchmark [4] released very recently. Other attempts to clas-

sify actions from soccer videos such as in [3] have focused on specific localization

tasks instead of classification. Therefore, we believe that our newly contributed

soccer dataset (SoccerDB1) and knowledge distillation based action recognition

framework (SoccerKDNet) would help progress the research on vision-based ac-

tion recognition in soccer videos.

In summary, we present SoccerDB1, a datset for action recognition in soc-

cer video. We also present SoccerKDNet, a knowledge distillation based action

recognition framework. We have achieved 67.20% accuracy on action recognition

task. Next, we describe our SoccerDB1 dataset in detail.

2

SoccerDB1 Dataset

We introduce a new soccer dataset named SoccerDB1 consisting of 448 soccer

video clips. The dataset contains videos of 4 action classes namely: Dribble, Kick,

Run, and Walk. There are over 70 video clips per class. Sample frames of differ-

ent action classes are shown in Figure 1. The video clips are created manually

from openly available broadcast soccer match videos available on YouTube. The

frames were sampled uniformly and each video clip contains 25-26 frames. The

proposed action recognition framework is discussed next.

Fig. 1. Sample frames from SoccerDB1 dataset, their actual and predicted labels.

SoccerKDNet

3

Fig. 2. SoccerKDNet: Schematic Architecture of Proposed End-to-End network.

3

Methodology

3.1

Knowledge Distillation based Transfer Learning

We propose the SoccerKDNet network to classify actions from soccer video clips

as shown in Figure 2. Here, we use the Temporal Adaptive Module (TAM) [9]

with backbones as both ResNet-50 and ResNet-101 [7]. We then add a few fully

connected layers with BatchNorm in front of the backbone network in order to

enable it to have some learnable parameters. The features from this setup are

then passed on to the frontnet which is shown in Figure 3. This entire setup is

referred to as the ’jointnet’ throughout the rest of the paper.

We use ResNet-18 as the student network in all our experiments. The ’joint-

net’ serves as the Teacher Network and is initially trained on the Soccer Dataset.

We use both ResNet-50 and ResNet-101 as backbones alongwith the Tempo-

ral Adaptive Module (TAM). We perform uniform sampling with all the video

frames in all our experiments since it is known to yield better results than dense

sampling [2].

In our network as shown in Figure 2, the various losses employed are described

below:-

– Cross Entropy Loss:-

The Cross Entropy Loss is given by the following formula:-

Lsoftmax = −

M

�

c=1

yo,c log(po,c)

(1)

– KullBack-Liebler Divergence Loss:-

The KullBack Liebler Divergence Loss can be given by the following formula:-

LKL =

M

�

c=1

ˆyc log ˆy**c

yc

(2)

We have also applied a Temperature (τ) hyperparameter to this equation.

4

Sarosij Bose et al.

– Knowledge Distillation Loss:-

If there is a given Teacher Network D and a Student Network S and the loss

of the student network is denoted by Lsoftmax and a hyperparameter α such

that (0 ≤ α ≤ 1), then the knowledge distillation Loss can be given by the

following formula:-

Lk.d. = α ∗ Lsoftmax + (1 − α) ∗ LKL

(3)

The yo,c represents the truth label for that particular sample, and po,c repre-

sents the softmax probabilities obtained after the final fully connected layer

(Fc4) L ∈ R128 x 4. Further, the ˆyc and yc are the predicted and the actual

probability distributions respectively for a given soccer frame sample.

Fig. 3. Architecture of the FrontNet Module. The Backbone network in Figure 2 has

an output of 400 classes so the obtained feature output serves as it’s input in this case.

3.2

Experiments

Datasets Used. Here, we use two datasets for our work. The first dataset is

the Kinetics 400 dataset which consists of 400 diverse action classes of everyday

activities with over 300, 000 videos. We used this dataset for pre-training our

backbone model as it can learn more generalized features.

For benchmarking our results, we used SoccerDB1. Further details regarding

our dataset have already been discussed in detail in Section 2.

Implementation Details. We first train the jointnet on the Soccer Dataset

for 100 epochs. The batch size was kept to 64 and the Cross Entropy Loss

function was used. The Adam optimizer was chosen with a learning rate of

0.0001 and CosineAnnealing rate decay scheduler.

Next, we train the student model which is a ResNet-18. We take a batch

size of 128 and train the student model for 200 epochs. The KullBack-Liebler

SoccerKDNet

5

Alpha (α) Model Accuracy

0.95

66.31%

0.90

67.20%

0.97

62.8%

Table 1. Alpha vs Model accuracy comparison

divergence loss function was used here alongwith the distillation loss as described

in equation 3. Table 1 illustrates the considerable change in accuracy with the

changing value of α, and we found the optimal value to be 0.90. The value

of Temperature (τ) was taken to be 6, SGD Optimizer was used here with

a constant learning rate of 0.0001, momentum of 0.9 and a constant weight

decay of 5e-4. Then the obtained student model was simply plugged into the

evaluation framework to obtain the corresponding accuracy. As evident from

Figure 4 in Section 4, experiments were run for 200 epochs till the validation

accuracy started saturating.

All the input video frames were resized to 224 × 224 RGB images using the

spatial cropping strategy as outlined in [2]. All the accuracies reported were

sampled over 5 runs to ensure the reproducibility of results. All models were

trained on a 32 GB NVIDIA V100 GPU.

4

Experimental Results

Accuracy Metrics. All accuracies reported here are Top-1 accuracies. All fig-

ures were sampled over 5 runs. The student model is ResNet-18 in all cases. In

our setting, we have used the Top-1/Top-5 accuracy metric for evaluation in all

our experiments as used in several previous works [6]. Top-1 accuracy refers to

when the 1st predicted model label matches with the ground truth label for the

particular frame. Using this metric, a particular soccer video is considered to be

classified correctly only when atleast half or more of it’s total frames match with

the corresponding ground truth label. Thus, we report the accuracies obtained

using various backbones in Table 2.

BackBone Teacher Acc. Student Acc.

ResNet-50

60.00%

65.26%

ResNet-101

62.10%

67.20%

Table 2. Top-1 validation accuracies obtained on the soccer dataset using various

network backbones.

We note that directly using the pre-trained backbone model yields a very

poor accuracy of 7.7% highlighting the need for a generalized network. We also

6

Sarosij Bose et al.

see that the student model, with proper training and sufficient number of epochs

exceeds the teacher model in accuracy.

When the fine-tuning dataset is small, it is very difficult to ensure the model

does not overfit to the dataset. Here, pre-training is carried out on the Kinetics

400 dataset, which is significantly larger in comparison to our Soccer Dataset. To

prevent overfitting, we rigorously apply regularizers such as Batch Normalization

on both the TAM and FrontNet module and dropout on the Student Model.

However, such concerns may still remain to some extent as highlighted in [1].

Model

Dataset Accuracy*

Russo et al. [10]

300

32.00%

Kukleva et al.† [8]

4152

94.50%

SoccerKDNet (R50)

448

65.26%

SoccerKDNet (R101)

448

67.20%

Table 3. Comparison of accuracy of SoccerKDNet with similar methods.

As it can be seen from Table 3, our model outperforms all other existing

models. The SoccerData dataset by Kukleva et al. [8] is an image dataset not

video dataset. On that particular dataset, the model requires digit level bounding

boxes and human keypoint annotations which our dataset does not have and

there are no trained models provided by the authors to be used publicly for

testing. Further, our models are trained on video data which cannot be ideally

tested on image datasets without compromising on crucial information such as

the temporal sequence present in a video.

Several earlier works such as Two Stream Networks [5] have millions of pa-

rameters and hence are unsuitable for edge deployment. Using knowledge dis-

tillation here, we show that even simple 2D networks such as ResNet-18 can be

used for action classification. Table 1 highlights this aspect by listing the net-

works used in our work - all of them have fewer than 50 mil. parameters and the

backbone network is frozen. For context, 3D networks such as C3D[15] has 73

million parameters.

From Figure 4, it can be seen that the student Model is able to achieve a Train

accuracy as high as 77.9% and validation accuracies above 60% underscoring

the generalizability and effectiveness of our proposed network. On the right, the

corresponding validation loss curve obtained using a ResNet-18 student and a

ResNet-50 teacher model is shown.

4.1

Ablation Study

We also perform a mini ablation study on a variety of factors: the backbone

network, stage in which distillation is applied, layers in frontnet module and

various hyperparameters.

SoccerKDNet

7

Fig. 4. On the left: red and blue curves denote the train and validation accuracies

respectively. On the right, the validation loss is shown.

– Backbone Network: As it can be seen from Table 2, we experimented with

2 backbone networks. The TAM-ResNet101 backbone model performs the

best for both the teacher and student models. Further, employing parameter

heavy models such as ResNet-152 as the backbone would not only defeat the

purpose of moving towards a more online solution but the relatively small

size of the target dataset for fine-tuning means there will be considerable

concerns in performance due to model over-fitting.

– Distillation Stage: There are two possibilities: directly distill the frozen TAM

backbone module and then use the distilled model as a plug-in within the

network. We call this as the early distillation where the distillation is done

early on in the network. However, performance using this approach is not

satisfactory: taking the TAM-ResNet50 as the backbone Teacher model, we

get only 8.3% accuracy on the Kinetics400 dataset using the ResNet-18 as

the student model. We suspect this is due to the inability of the student

model to directly learn features from the heavy teacher model. Therefore,

we chose to go with the late stage distillation process.

– FrontNet Module and Hyperparameters: We found adding dropout layers

decreased the accuracy by 1.2%, adding more fully connected layers in the

latter half of the FrontNet module decreased accuracy as much as 4.9% of

the teacher module to 57.2%. We did not find much difference in accuracy

on using NLL loss instead of CrossEntropy so we chose to keep it. We found

a learning rate of 0.0001 and batch size of 64 and 128 to be optimal for all

our experiments as bigger batches lead to better performance.

5

Conclusion

In this paper, we introduce a new Soccer Dataset consisting of 4 diverse classes

and over 70 video clips per class. The network not only provides the flexibility of

using any of the original (teacher) model as the classifier but also has the option

of using a smaller network (ResNet-18 in our case) as the student network. In

future, we plan to add more action classes in our dataset. Also we plan to utilize

SoccerKDNet for soccer event detection based on the actions of the players.

8

Sarosij Bose et al.

References

1. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the

kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and

Pattern Recognition. pp. 6299–6308 (2017)

2. Chen, C.F.R., Panda, R., Ramakrishnan, K., Feris, R., Cohn, J., Oliva, A., Fan, Q.:

Deep analysis of cnn-based spatio-temporal representations for action recognition.

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern

Recognition. pp. 6165–6175 (2021)

3. Cioppa, A., Deliege, A., Giancola, S., Ghanem, B., Droogenbroeck, M.V., Gade,

R., Moeslund, T.B.: A context-aware loss function for action spotting in soccer

videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and

Pattern Recognition. pp. 13126–13136 (2020)

4. Deliege, A., Cioppa, A., Giancola, S., Seikavandi, M.J., Dueholm, J.V., Nasrollahi,

K., Ghanem, B., Moeslund, T.B., Van Droogenbroeck, M.: Soccernet-v2: A dataset

and benchmarks for holistic understanding of broadcast soccer videos. In: Proceed-

ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

pp. 4508–4519 (2021)

5. Feichtenhofer, C., Pinz, A., Zisserman, A.: Convolutional two-stream network fu-

sion for video action recognition. In: Proceedings of the IEEE conference on com-

puter vision and pattern recognition. pp. 1933–1941 (2016)

6. Girdhar, R., Tran, D., Torresani, L., Ramanan, D.: Distinit: Learning video rep-

resentations without a single labeled video. In: Proceedings of the IEEE/CVF

International Conference on Computer Vision. pp. 852–861 (2019)

7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In:

Proceedings of the IEEE conference on computer vision and pattern recognition.

pp. 770–778 (2016)

8. Kukleva, A., Khan, M.A., Farazi, H., Behnke, S.: Utilizing temporal information

in deep convolutional network for efficient soccer ball detection and tracking. In:

Robot World Cup. pp. 112–125. Springer (2019)

9. Liu, Z., Wang, L., Wu, W., Qian, C., Lu, T.: Tam: Temporal adaptive module for

video recognition. In: Proceedings of the IEEE/CVF International Conference on

Computer Vision. pp. 13708–13718 (2021)

10. Russo, M.A., Filonenko, A., Jo, K.H.: Sports classification in sequential frames

using cnn and rnn. In: 2018 International Conference on Information and Commu-

nication Technology Robotics (ICT-ROBOT). pp. 1–3. IEEE (2018)

11. Sarkar, S., Chakrabarti, A., Mukherjee, D.P.: Estimation of ball possession statis-

tics in soccer video. In: Proceedings of the 11th Indian Conference on Computer

Vision, Graphics and Image Processing. pp. 1–8 (2018)

12. Sarkar, S., Mukherjee, D.P., Chakrabarti, A.: From soccer video to ball possession

statistics. Pattern Recognition 122, 108338 (2022)

13. Sarkar, S., Mukherjee, D.P., Chakrabarti, A.: Watch and act: Dual interacting

agents for automatic generation of possession statistics in soccer. In: Proceedings

of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp.

3560–3568 (2022)

14. Sarkar, S., Mukherjee, D.P., Chakrabarti, A.: Reinforcement learning for pass de-

tection and generation of possession statistics in soccer. IEEE Transactions on

Cognitive and Developmental Systems 15(2), 914–924 (2023)

15. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotem-

poral features with 3d convolutional networks. In: Proceedings of the IEEE inter-

national conference on computer vision. pp. 4489–4497 (2015)

1. 网页内容摘要

该网页详细介绍了YouTube关于第三方AI训练的政策。YouTube允许创作者和版权持有者选择是否允许第三方AI公司使用其内容进行AI模型训练。默认情况下，第三方训练设置是关闭的，用户无需采取任何行动即可保护其内容不被用于第三方训练。只有当内容符合特定条件时才有资格用于第三方训练：相关权利持有者允许、视频为公开状态且符合YouTube服务条款和社区准则。

创作者可以选择允许特定的第三方公司或所有第三方公司使用其内容。一旦开启此设置，YouTube可能会与被允许的第三方公司分享视频内容，用于训练AI模型。视频的训练许可状态将通过YouTube Data API公开可见。用户可以随时在YouTube Studio或YouTube Studio Content Manager中更改第三方训练设置，但变更可能需要最多7天才能在公开接口中反映。

网页还包含常见问题解答，涉及如何更改设置、第三方公司如何使用内容、第三方公司如何加入列表以及是否会因内容被用于训练而获得报酬等问题。

2. 与探究问题相关的信息提取

YouTube第三方训练政策的核心内容和授权机制如下：

1. 政策基本原则：

   • YouTube承诺负责任地开发AI，确保创作者和权利持有者能够控制其内容的使用方式
   • 默认情况下，第三方训练设置是关闭的，用户无需采取行动即可保护内容
   • YouTube的服务条款明确禁止未经授权的内容使用，如未授权下载和抓取

2. 内容资格条件：
   内容必须同时满足以下条件才有资格用于第三方训练：

   • 视频的相关权利持有者允许第三方训练
   • 视频隐私设置为"公开"
   • 视频符合YouTube的服务条款和社区准则

3. 授权机制详情：

   • 创作者/权利持有者可以选择开启第三方训练设置
   • 可以从列表中选择特定第三方公司，或选择允许所有第三方公司
   • 对于使用YouTube Studio的创作者：授权内容包括上传到其频道的所有符合条件的视频
   • 对于使用YouTube Studio Content Manager的权利持有者：授权内容包括所有符合条件的已认领视频和关联频道中的视频
   • 视频的训练许可状态将通过YouTube Data API公开可见
   • 只有当所有相关权利持有者都允许某公司时，YouTube才会与该第三方公司分享视频

4. 设置管理：

   • 频道所有者可以在YouTube Studio的频道设置中查看和更改第三方训练设置
   • 管理员可以在YouTube Studio Content Manager的设置页面中查看和更改第三方训练设置
   • 设置可以随时更改，但变更可能需要最多7天才能在公开接口(YouTube Data API)中反映

5. 重要注意事项：

   • 即使创作者允许，第三方公司也可能不会最终使用其内容
   • YouTube无法控制第三方公司的行为，因为它们独立于YouTube/Google
   • 选择"所有第三方公司"意味着允许任何第三方公司，即使它们未在列表中
   • 第三方公司如希望加入列表，可以申请加入
   • 目前YouTube不促成第三方公司与创作者或其他权利持有者之间的付款

这一政策体现了YouTube在AI发展中平衡创新与内容创作者权益的努力，通过明确的选择机制和透明度，确保创作者对其内容的控制权。

3. 相关网页链接

1. 负责任的AI开发 - 上下文：YouTube承诺负责任地开发AI，确保创作者和权利持有者继续控制其内容的使用方式

2. 生成式AI - 上下文：一些创作者和权利持有者可能希望与AI公司合作，帮助开发最新的生成式AI技术

3. 服务条款 - 上下文：YouTube的服务条款禁止未经授权使用内容，如未经授权的下载和抓取

4. 公开 - 上下文：视频隐私设置必须为公开才有资格用于第三方训练

5. 社区准则 - 上下文：视频必须符合YouTube的社区准则才有资格用于第三方训练

6. 了解更多 - 上下文：关于权利持有者使用YouTube Studio Content Manager的内容包括所有符合条件的已认领视频和关联频道中的视频的更多信息

7. 频道设置 - 上下文：频道所有者可以在YouTube Studio的频道设置中查看和更改第三方训练设置

8. 设置 - 上下文：管理员可以在YouTube Studio Content Manager的设置页面中查看和更改第三方训练设置

9. 申请 - 上下文：希望与YouTube创作者和权利持有者合作训练AI模型的第三方公司可以申请加入列表

4. 相关图片

网页中没有与YouTube第三方训练政策相关的图片。

5. 多语言支持

我已根据您的输入语言(中文)提供了相应的中文回复。如果您希望使用其他支持的语言(英文、日语、韩语、繁体中文、西班牙语、葡萄牙语)，请告知我，我可以相应调整。

摘要

YouTube正在推出一项新功能，允许创作者选择是否让第三方公司使用其视频来训练AI模型。该功能默认关闭，意味着创作者无需采取任何行动即可阻止第三方抓取其视频用于AI训练。YouTube团队成员Rob表示，这是"支持创作者并帮助他们在AI时代从YouTube内容中获取新价值的重要第一步"。该设置将在未来几天内在YouTube Studio中推出，未经授权的抓取仍然被禁止。创作者可以从指定的第三方公司列表中选择，或允许所有第三方公司训练其视频。初始合作公司名单包括AI21 Labs、Adobe、Amazon、Anthropic、Apple等17家科技公司。值得注意的是，Google自身已经使用YouTube数据来训练其AI工具，而此前有报道称OpenAI、Apple和Anthropic等公司的AI模型也使用了从YouTube抓取的内容和数据集进行训练。

与问题相关的信息提取

关于YouTube第三方训练政策的发布时间和具体内容，网页提供了以下信息：

发布时间：

• 文章发布于2024年12月17日
• YouTube将在"未来几天内"在YouTube Studio中推出这项设置功能
• 该功能最初在2024年9月被宣布正在开发中

政策具体内容：

1. YouTube推出了允许创作者选择是否让第三方公司使用其视频来训练AI模型的功能
2. 该功能默认关闭，意味着创作者如不希望第三方公司抓取其视频用于AI训练，无需采取任何行动
3. 创作者可以从指定的第三方公司列表中选择特定公司，或允许所有第三方公司训练其视频
4. 初始合作的第三方公司名单包括17家科技公司：AI21 Labs、Adobe、Amazon、Anthropic、Apple、ByteDance、Cohere、IBM、Meta、Microsoft、Nvidia、OpenAI、Perplexity、Pika Labs、Runway、Stability AI和xAI
5. YouTube发言人Jack Malon表示，这些公司"被选中是因为它们正在构建生成式AI模型，并且可能是与创作者潜在合作的合理选择"
6. 未经授权的抓取仍然被禁止
7. 值得注意的是，Google自身已经使用YouTube数据来训练其AI工具，Google在9月表示："多年来，我们一直使用上传到YouTube的内容来改善创作者和观众在YouTube和Google上的产品体验，包括通过机器学习和AI应用。我们这样做符合创作者同意的条款。"

相关网页链接

1. YouTube支持帖子链接 - 上下文：TeamYouTube成员Rob发布的关于第三方AI可训练性的支持帖子

2. YouTube支持页面 - 上下文：详细说明创作者如何选择第三方公司以及可以如何更改设置

3. TechCrunch报道链接 - 上下文：提供了初始第三方公司列表的报道

4. OpenAI使用YouTube数据的报道 - 上下文：关于OpenAI使用YouTube内容训练其模型的报道

5. Apple和Anthropic使用YouTube数据的报道 - 上下文：关于Apple和Anthropic使用YouTube视频作为训练数据的报道

6. YouTube 9月公告 - 上下文：YouTube在9月宣布该功能正在开发中的博客文章

相关图片

1. 图片标题：YouTube logo image in red over a geometric red, black, and cream background
   内容：YouTube标志图像，红色标志放置在几何形状的红色、黑色和米色背景上
   来源：Alex Castro / The Verge
   链接：https://platform.theverge.com/wp-content/uploads/sites/2/chorus/uploads/chorus_asset/file/23986639/acastro_STK092_03.jpg?quality=90&strip=all&crop=0%2C0%2C100%2C100&w=2400

2. 图片标题：Jay Peters
   内容：Jay Peters的头像照片，文章作者
   来源：The Verge
   链接：https://platform.theverge.com/wp-content/uploads/sites/2/chorus/author_profile_images/195819/JAY_PETERS.0.jpg?quality=90&strip=all&crop=0%2C0%2C100%2C100&w=96