A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Research of efficiency improvements of the slowfast action recognition architecture ; „SlowFast“ veiklos atpažinimo architektūros efektyvumo didinimo tyrimas
This bachelor thesis explores the task of action recognition from videos and possible methods to make the SlowFast [CHJ+19] architecture, which achieves state-of-the-art results in this field, more lightweight and efficient. The first part of this work introduces the most noteworthy past methods and approaches as well as the most popular datasets used for training and testing. Next, the SlowFast [CHJ+19] neural network is introduced: the main idea, example architecture and achieved results. In the following section, each of the performed experiments is presented, which can be separated into two main categories: changes made to the architecture of the network itself and modifications performed on the input frames of the videos. Changes to architecture consisted of reducing the number of channels and modifying the ResNet-50 backbone. The input was experimented with by reducing frame and crop size as well as reducing the number of frames passed. Afterwards, the accuracies and other information, such as used memory resources and time required to train the network, of all of modifications are summarized and compared with. There were modifications from both categories, that decreased used resources and training time without significantly losing accuracy, but overall, the networks with changed input trained faster, and used fewer resources to achieve the same or slightly better accuracy. The only disadvantage of them compared to the modifications with changed architecture is that they do not reduce the model size and the number of parameters.
Research of efficiency improvements of the slowfast action recognition architecture ; „SlowFast“ veiklos atpažinimo architektūros efektyvumo didinimo tyrimas
This bachelor thesis explores the task of action recognition from videos and possible methods to make the SlowFast [CHJ+19] architecture, which achieves state-of-the-art results in this field, more lightweight and efficient. The first part of this work introduces the most noteworthy past methods and approaches as well as the most popular datasets used for training and testing. Next, the SlowFast [CHJ+19] neural network is introduced: the main idea, example architecture and achieved results. In the following section, each of the performed experiments is presented, which can be separated into two main categories: changes made to the architecture of the network itself and modifications performed on the input frames of the videos. Changes to architecture consisted of reducing the number of channels and modifying the ResNet-50 backbone. The input was experimented with by reducing frame and crop size as well as reducing the number of frames passed. Afterwards, the accuracies and other information, such as used memory resources and time required to train the network, of all of modifications are summarized and compared with. There were modifications from both categories, that decreased used resources and training time without significantly losing accuracy, but overall, the networks with changed input trained faster, and used fewer resources to achieve the same or slightly better accuracy. The only disadvantage of them compared to the modifications with changed architecture is that they do not reduce the model size and the number of parameters.
Research of efficiency improvements of the slowfast action recognition architecture ; „SlowFast“ veiklos atpažinimo architektūros efektyvumo didinimo tyrimas
Kavaliauskas, Svajūnas (author) / Grevys, Andrius
2020-06-09
Theses
Electronic Resource
Lithuanian , English
DDC:
720
Gaisro miškuose atpažinimo sistemos tyrimas / ; Research of forest fire detection system.
| BASE | 2022
Saulės baterijų efektyvumo tyrimas orlaivyje / ; Solar cell efficiency research on aerial vehicle.
| BASE | 2018