TAMFORMER MULTI-MODAL TRANSFORMER WITH LEARNED ATTENTION MASK FOR EARLY INTENT PREDICTION Nada Osman Guglielmo Camporese Lamberto Ballan

2025-05-02 0 0 2.68MB 5 页 10玖币

侵权投诉

TAMFORMER: MULTI-MODAL TRANSFORMER WITH LEARNED ATTENTION MASK

FOR EARLY INTENT PREDICTION

Nada Osman, Guglielmo Camporese, Lamberto Ballan

Department of Mathematics “Tullio Levi-Civita”, University of Padova, Italy

{nadasalahmahmoud.osman, guglielmo.camporese}@phd.unipd.it

lamberto.ballan@unipd.it

ABSTRACT

Human intention prediction is a growing area of research

where an activity in a video has to be anticipated by a vision-

based system. To this end, the model creates a representation

of the past, and subsequently, it produces future hypothe-

ses about upcoming scenarios. In this work, we focus on

pedestrians’ early intention prediction in which, from a cur-

rent observation of an urban scene, the model predicts the

future activity of pedestrians that approach the street. Our

method is based on a multi-modal transformer that encodes

past observations and produces multiple predictions at dif-

ferent anticipation times. Moreover, we propose to learn the

attention masks of our transformer-based model (Temporal

Adaptive Mask Transformer) in order to weigh differently

present and past temporal dependencies. We investigate our

method on several public benchmarks for early intention pre-

diction, improving the prediction performances at different

anticipation times compared to the previous works.

Index Terms—Action anticipation, multi-modal deep learn-

ing, transformers, pedestrian intent prediction

1. INTRODUCTION

In the last years, the development of computer vision algo-

rithms has seen a massive improvement thanks to the advent

of deep learning enabling new applications in the context of

autonomous driving, video surveillance, and virtual reality.

The visual understanding capabilities of deep learning models

have been adopted in various domains, from smart cameras

used in video surveillance to cognitive systems in robotics

and multi-modal sensors for autonomous driving. Moreover,

a recent interesting direction involves predicting future ac-

tivities that can be anticipated from a visual content [1,2,

3]. Some applications enabled by the models designed for

action anticipation are pedestrian intention prediction from

a smart camera and ego-centric action anticipation from a

robotic agent. In this work, we investigate the early inten-

tion prediction of pedestrians in an urban environment. In

particular, i) we propose a new model for early intent pre-

diction based on a multi-modal transformer; ii) we propose

a new mechanism for learning the attention masks inside the

transformer that leads to better performances and more efﬁ-

cient computation; and iii) we conduct several experiments

and model ablations on different datasets obtaining state-of-

the-art results on the early intent prediction task.

2. RELATED WORKS

Action Recognition. Video action recognition is a well inves-

tigated problem that, in recent years, has experienced massive

improvements thanks to the recent progress of deep learning.

Speciﬁcally, traditional hand-crafted video approaches [4,5,

6,7,8] have been replaced by models based on recurrent neu-

ral networks [9,10,11,12,13], 2D CNN [14,15,16], and 3D

CNN [17,18,19,20,12,21]. Transformers [22] have been

also investigated for spatio-temporal modeling [23,24,20]

improving the state-of-the-art performances on video related

problems including video action recognition.

Action Anticipation and Intent Prediction. Recently, antic-

ipating actions on videos gained attention given the develop-

ment of new methods [1,25,3,26,27], datasets [28,29,30],

and applications such as autonomous driving, human-robot

interaction and virtual reality. In particular, in urban environ-

ments, the pedestrian intent prediction from third-view cam-

eras is a growing area [29,30] in which models are designed

to predict the future activity of pedestrians.

Temporal Modeling on Vision Problems. Video-based

models need to process spatial and temporal information.

Usually, the temporal axis is considered an independent com-

ponent of the video, and in the model design, the spatial

information is processed differently from the temporal one.

Recent works proposed to model the temporal at different

frame rates [19], with multiple time-scales [25,31], and with

an adaptive frame-rate [32]. However, in such works, the

temporal sampling strategy of the frames is ﬁxed and treated

as a hyper-parameter of the model. For this reason, in this

work, we explore and propose an adaptive mechanism for

weighting the importance of the current and past frames by

learning the attention mask inside our transformer model.

arXiv:2210.14714v1 [cs.CV] 26 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TAMFORMER:MULTI-MODALTRANSFORMERWITHLEARNEDATTENTIONMASKFOREARLYINTENTPREDICTIONNadaOsman,GuglielmoCamporese,LambertoBallanDepartmentofMathematicsTullioLevi-Civita,UniversityofPadova,Italyfnadasalahmahmoud.osman,guglielmo.camporeseg@phd.unipd.itlamberto.ballan@unipd.itABSTRACTHumanintentionpredict...

展开>> 收起<<

TAMFORMER MULTI-MODAL TRANSFORMER WITH LEARNED ATTENTION MASK FOR EARLY INTENT PREDICTION Nada Osman Guglielmo Camporese Lamberto Ballan.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

TAMFORMER MULTI-MODAL TRANSFORMER WITH LEARNED ATTENTION MASK FOR EARLY INTENT PREDICTION Nada Osman Guglielmo Camporese Lamberto Ballan

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: