Difference between revisions of "Private:progress-diab"
From NMSL
Line 3: | Line 3: | ||
=== May 18 === | === May 18 === | ||
+ | * Storage management | ||
+ | |||
+ | * Measurement study | ||
+ | |||
+ | * FVV streaming | ||
= Spring 2014 (RA) = | = Spring 2014 (RA) = |
Revision as of 20:40, 18 May 2014
Summer 2014 (RA)
- Courses: NA
May 18
- Storage management
- Measurement study
- FVV streaming
Spring 2014 (RA)
- Courses:
- CMPT-705: Design and Analysis of Algorithms
- CMPT-820: Multimedia Systems
- Submissions:
- NA
- Publications:
- [accepted] K. Diab, T. ElGamal, K. Calagary, and M. Hefeeda, "Storage optimization for 3d streaming systems," in Proc.of ACM Multimedia Systems Conference (MMSys’14), Singapore, March 2014.
- Submissions Plans:
- ACM Multimedia in March 31, 2014 (http://www.acmmm.org/2014/)
Mar 23
- FVV project:
- Implementation of a program to calculate PSNR and MSE values for reference views between the uncompressed videos and the 10 representations we have. I wrote a script to automate the calculations over our dataset.
- For the next 3-5 days, I will work on 'related work' section and start the evaluation plan.
- Courses work:
- Algorithm: quiz, I spent some time reading from the textbook and external materials.
- Multimedia: paper presentation on perceptual issues in stereoscopic images, the presentation lasted 80 min divided on two classes; there was an interest from the audience to discuss different aspects of the paper. Also, I proposed an idea for the term project, where I am currently working on.
Feb 23
- No major progress in the research of the FVV project. Mainly, this is because I was investigating different ideas to control the segment buffer.
- Some extensions ideas had been discussed with Ahmed Hamza and I added them to my notes.
- In the implementation part, I was reading the libdash code and the current implementation of Ahmed Hamza. I think that I had discovered the reason of the unexpected behavior in case of view switching:
- Segment buffers are controlled by one thread. Each frame buffer is controlled by a different thread. Also, there is a thread to handle the view switching. When the view switching happens, the view switching thread tries to modify the buffers while other threads accessing these buffers. There are locks, and here comes the problem of the deadlock.
- My solution is: when the view switching happens, the view switching thread sends a signal/message to other threads to modify their buffers. The view switching thread mustn't modify the buffers, hence, avoiding deadlocks.
- I discussed the solution with Ahmed Hamza and it looks fine, I will applying it to the current implementation and see what happens.
- For the courses:
- Multimedia: I will deliver the H264 encoder/decoder this week. Also, I will present the first paper.
- Algorithms: I have a quiz in Feb. 25, 2014.
Feb 07
- Currently, I am working with Ahmed Hamza for the FFV rate adaptation project
- Working, with Ahmed, on the optimization problem formulation
- Specifically, I am investigating to employ the queuing theory in the formulation to increase decoders utilization and keep buffers full as we can. We model each decoder as a server, and each segment as a client. Some challenges are modeling the arrival rate distribution, and relating the queuing theory formulas to our formulas.
- Number of decoders may increase in order to increase the throughput of the whole decoding step.
- Working on re-implementing this pipeline in the client-side: downloading -> buffering -> decoding. Some of our objectives:
- Reduce overheads of spawning new threads for decoders and other steps. Each group of threads should be in a thread pool.
- Increase the decoding throughput. Currently, the segments at time "t" are decoded in order, even if there are some downloaded segments of time "t+1". Mainly, we need to keep in-order decoding but enable decoding of segments at different timestamps. To do this, I will remove the decoding thread (which acts as a distributor but a barrier at the same time) and use a push model to push segments to decoders pool thread. On the output side, I will use a priority queue rather than the normal queue, to store the decoded frames in-order. Frames should be indexed properly.
- The current control flow is not easy to change or modify which may lead to more bugs. I am trying to simplify and organize the control flow. For example, what does the player do when the user changes the viewpoint?
- There was a discussion about moving some of the processing to the server side.
Jan 24
- I worked on the camera ready version of our MMSys'14 paper.
- Currently, I am working on the literature survey for the rate adaptation project. This report should have been delivered tonight, but I couldn't complete all the sections that I want to write. I will deliver it and complete it by the next weekend.
- For the rate adaptation project:
- There had been different approaches in the academia for 2D videos.
- Rate adaptation for 3D videos are rare. This means -somehow- that we are in the good direction.
- In the next four weeks, I should take some important decisions. Some questions that I need to answer:
- Do we need to deal with 3D videos as two different views? and not only as a 2D image. This will let us exploit some facts about 3D videos.
- There are many objectives that define a "good" rate adaptation algorithms. What objectives do we need to handle in our algorithm?
- What is the problem definition?
- How are we going to model the problem and derive the solution? I am thinking of two approaches right now. Optimization problem or Vector space modeling. In both cases, a "good" objective function or a "good" similarity metric should be designed.
- Human Visual System for 3D perception should be included in the model as well
- How to evaluate our model? (Simulation, Emulation, real system?)
- How to integrate such a model with the current system? How to extend the current storage optimization algorithm to support the storage requirements of multiple-bitrate versions?
- For the courses:
- Multimedia: I will deliver a hierarchical JPEG encoder/decoder implementation next week. I had implemented it using OpenCV. TODO components are: Huffman coding (can use external code), hierarchical flow, GUI and the document. Also, I will present two papers (during the course):
- 3D-TV content creation: Automatic 2D-to-3D video conversion
- Perceptual issues in stereoscopic signal processing
- Algorithms: I have a quiz in Feb. 4, 2014.
- Multimedia: I will deliver a hierarchical JPEG encoder/decoder implementation next week. I had implemented it using OpenCV. TODO components are: Huffman coding (can use external code), hierarchical flow, GUI and the document. Also, I will present two papers (during the course):