Chris Pollett >
Students >
Lei
( Print View)
[Bio]
[Blog]
[CS 297 Proposal]
[GANs-PDF]
[Gen Videos Scene Dynamics-PDF]
[3D CNNs Human Action RECOG-PDF]
[TGAN-PDF]
[Deliverable 1]
[Deliverable 2]
[Deliverable 3]
[Deliverable 4]
[CS297 Report - PDF]
[CS 298 Proposal]
[CS 298 Report - PDF]
[CS 298 Slides - PDF]
|
Project Blog
Week 8- Mar 24, 2020:
What was done:
- Generated 11 different emotions directions with a labeled dataset
- Generated videos with random human face and emotions in the latent space
What to do:
- Add more emotions
- Try to improve the generated videos: 1) less hair moves 2) No face changes
- Training videos and predict the emotions in a sequence
Week 7- Mar 17, 2020:
What was done:
- Use stylegan2 latent space
- Improved the video pose transfer to stylegan2 latent space
What to do:
- Generate different emotions in the latent space
Week 6- Mar 10, 2020:
What was done:
- Video pose transfer
- Predict noise vector directly from image into latent space
What to do:
- Video generation in the latent space
- Optimize the video pose translation
Week 5- Mar 3, 2020:
What was done:
- Used LSTM to predict latent vectors
- Generate intepolate video directly from latent space
What to do:
- Take a video of different face poses and transfer the style to random face
- Find a longer video for LSTM prediction training
Week 4- Feb 25, 2020:
What was done:
- Training with Progressive Growing of GANs for UCF101 dataset.
- Tried to recover human face image from pre-trained latent space with Flickr-Faces-HQ dataset.
- Explored the trained latent space
What to do:
- Apply to videos for searching a pre-trained GAN latent space
Week 3- Feb 18, 2020:
What was done:
- Implemented a 2D + 1D GAN to generate videos.
- Use LSTM to generate 16 noise vectors, and use them to generate 16 frames.
- Cannot directly train a pre-trained generator due to backpropagation problem.
What to do:
- One way to use a pre-trained GAN latent space
- Create a traing dataset from recovered latent space, which are image to noise vector mappings
- Train a GAN to find the closest mappings from giving frames to noise vectors
- Recover a synthesis video from the noise vectors
Week 2- Feb 11, 2020:
What was done:
- Implemented paper 2. Created a short video by search pre-trained GAN latent space.
What to do:
- Write a GAN model to generate a latent vector to mapping to its original image. Write a model to predict next frames.
Week 1- Feb 4, 2020:
What was done:
- Finished the CS298 proposal.
What to do:
- Implement paper 2, and search GAN latent space.
Week 15- Dec 3, 2019:
What was done:
- Finished the second version of CS297 report.
What to do:
- Finish the final changes of CS297 report and submit to canvas. Fix w3c html errors.
Week 14- Dec 3, 2019:
What was done:
- Finished draft version of CS297 report.
What to do:
- Some changes on the draft report.
Week 13- Nov 26, 2019:
What was done:
- Finished deliverable 4. Created a few slides for video generation idea.
What to do:
- Draft version of CS297 report.
Week 12- Nov 19, 2019:
What was done:
- Use lstm layer to simplify 3D GAN to 2D.
What to do:
- Finish deliverable 4. Create a few slides for video generation idea.
Week 11- Nov 12, 2019:
What was done:
- Read paper 8 and tried to improve video quality with a separate temporal layer. Tried to generate video with 8 frames which doesn't work.
What to do:
- Refine the video generation model.
Week 11- Nov 05, 2019:
What was done:
- Futher improved quality of generated videos with pix2pix technologies.
What to do:
- Trying to build a hypothesis to use tiered 2D gan plus 1D temporal layer and temporal layer to generate fake videos instead of a 3D gan.
- Try to keep investigating pix2pix to video method
Week 10- Oct 29, 2019:
What was done:
- Improved the video qualities with labeld images and pix2pix framework. Able to generate moving objects but they are not clear.
What to do:
- Try to further improve video quality with pix2pix framework. Use one frame to many frames mappings when training.
Week 9- Oct 22, 2019:
What was done:
- 3D gan seems not working well to generate fake videos. I got only static background after training 3D video gan for a day.
What to do:
- Try to explore labeled technologies to reduce the complexity of images
Week 8- Oct 15, 2019:
What was done:
- Create a model to generate videos
What to do:
- Improve the quality of generated videos
Week 6- Oct 1, 2019:
What was done:
- Slides for 3D Convolutional Neural Networks for Human Action Recognition
What to do:
- Finish deliverable 2
Week 5- Sep 24, 2019:
What was done:
- Improved deliverable 1
- Slides for Generating Videos with Scene Dynamics
What to do:
- Read paper: 3D Convolutional Neural Networks for Human Action Recognition and write slides
Week 4- Sep 17, 2019:
What was done:
- Finished deliverable 1
- Modified project proposal, add bio info, and add blogs
What to do:
- Make improvement of GAN training dataset
- Upload deliverable 1
- Read paper: Generating Videos with Scene Dynamics and write slides
Week 3- Sep 10, 2019:
What was done:
- Wrote slides for paper 7
- Read book: GANs in Action
What to do:
- Write a GAN program to fake Chinese numbers
- Modify project proposal, add bio info, and add blogs
- Check paper Eigenheads for Reconstruction
Week 2- Sep 3, 2019:
What was done:
- Read paper: Generative Adversarial Nets
- Write project proposal
What to do:
- Write a GAN program to fake Chinese numbers
- Write slides for paper 7
Week 1- Aug 27, 2019:
What was done:
- Discussed projects.
What to do:
- Write project proposal
|