Seattle, Washington, United States
4K followers 500+ connections

Join to view profile

About

A system backend software engineer with expertise from SW to HW, especially regarding…

Activity

Join now to see all activity

Experience & Education

  • Meta

View YU-CHENG’s full experience

See their title, tenure and more.

or

By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.

Licenses & Certifications

  • Neural Networks for Machine Learning Graphic

    Neural Networks for Machine Learning

    Coursera

    Issued

Volunteer Experience

Projects

  • [ECS 289G Scalable Machine Learning] Implement and compare performance of different neural network architecture for first person video recognition

    As the first-person cameras like GoPro become popular for some sports hobbyists and geeks, there would be more first-person videos in the near future. Most of the video online, such as youtube, are shot in third person video. And, the research of video action recognition on first-person videos is still in the beginning. We want to see if the deep learning models can also apply on the recognition of the first-person videos and whether they perform well as for the images and third-person videos…

    As the first-person cameras like GoPro become popular for some sports hobbyists and geeks, there would be more first-person videos in the near future. Most of the video online, such as youtube, are shot in third person video. And, the research of video action recognition on first-person videos is still in the beginning. We want to see if the deep learning models can also apply on the recognition of the first-person videos and whether they perform well as for the images and third-person videos. At first, we did experiments on the popular deep CNN model, Caffe’s Imagenet pre-train model, which adopts the same architecture as AlexNet. Based on this model, we fine-tune AlexNet by training both first-person and third-person point-of- view frames from videos. Then we use un-training dataset to evaluate classification accuracy rate of first-person and third-person point-of-view videos. Besides, to find a better network for recognizing first-person video, we also fine-tune C3D and GoogLeNet on Caffe. Results of three networks for C3D, Alexnet and GoogleNet are 76%, 68% and 68% separately. We can see that C3D gets the best performance among these three architectures for video recognition. This suggests that including motion information helps improve video accuracy rate.

    Other creators

  • [ECS 289G Visual Recognition] Pair Selection for View Invariance in Siamese Network

    -

    First-person cameras are getting prevalent for sport enthusiasts, but the amount is still quite few. At the same time, convolutional neural networks (CNNs) made a big step forward since the occurrence of Alexnet. To take advantage of the existing numerous third-person videos on Internet and the success of CNNs, the transfer learning with the help of the Siamese network is a promising way to raise the video classification accuracy of first-person videos, given limited first-person video training…

    First-person cameras are getting prevalent for sport enthusiasts, but the amount is still quite few. At the same time, convolutional neural networks (CNNs) made a big step forward since the occurrence of Alexnet. To take advantage of the existing numerous third-person videos on Internet and the success of CNNs, the transfer learning with the help of the Siamese network is a promising way to raise the video classification accuracy of first-person videos, given limited first-person video training dataset.

    To find the mapping between different views, we use the dataset that has synchronized videos in first-person view and third-person views. However, when the pair amount becomes larger, the training time takes longer. Therefore, in this work, we try to find the best strategy to do pair selection for view invariance training in Siamese network. From our experiment results, we show that pair selection should focus on hard negatives and clarify positive boundary by compensation. In addition, choosing an appropriate feature is very important to gain better results for view invariance training in Siamese network for multiple view videos.

    Other creators
    See project

More activity by YU-CHENG

View YU-CHENG’s full profile

  • See who you know in common
  • Get introduced
  • Contact YU-CHENG directly
Join to view full profile

Other similar profiles

Explore top content on LinkedIn

Find curated posts and insights for relevant topics all in one place.

View top content

Others named YU-CHENG LIN in United States

21 others named YU-CHENG LIN in United States are on LinkedIn

See others named YU-CHENG LIN

Add new skills with these courses

See all courses