Prasad B @Dolby Laboratories, SF: Comments and reviews.

This post contains edited, non-confidential feedback given by Prasad Balasubramanian for the answers received by him on GapJumpers.


Question:

Design and implement a layered depth of field blur based on Gaussian Blur.

Feedback:

I am looking for an implementation of a layered depth of field blur on an image based on the knowledge of it’s corresponding depth value. If depth values are in the range of 0…1 (32-bit float), you could assume that an image pixel that is closest to you (depth of 1.0) is in focus. You may then define a set of ‘10' blur kernels (2x2, 3x3, 4x4 … ) at a step of 0.1, say for instance, pick the kernel for a image pixel based on the corresponding depth value.
Depth-of-field blur is used in 2D content to give the impression of 3D. It is a fairly common filter applied to the main camera in games and uses the z(depth buffer) to blur objects that are distant from focal point.

Question:

Study approaches to automatic logo detection in broadcast video and propose strategies for real-time GPU implementation.
Prepare a research report that compares existing approaches. Evaluate feasibility of real-time implementation on a GPU.

Feedback:

I am interested in your thoughts on how you could take advantage of the temporal redundancy inherent in video applications to track logos once detected; and how this approach can be extended to further improve accuracy of detection. What are it’s likely implications on GPU video/ texture memory requirements and performance?
Would you also consider implementing scale space computation of uncompressed HD Video using OpenCL or Cuda ?

(Answer questions on GapJumpers and receive top quality feedback from industry experts. The complete feedback is confidential only to the applicants.)

Email me when GapJumpers publishes or recommends stories