Yesterday: (last couple of weeks) Adapted Perera et al.'s (2012) datasets for input into my own system. Included changes for thresholding. Checks on timing and accuracy, though not well validated yet. Will need to adjust to have solid accuracy checking. Broad overview of polynomial time algorithms for maximal cliques.

Good discussion to lay groundwork for multi-resolution processing.

3D Adjacency
Data from Perera et al.
(2012) 3books

2D View
Data from Perera et al.
(2012) 3books

Today: Wrapped up Perera et al. dataset adaption. After finding a bug in randomized ordering to ensure input order is inconsequential, I have success!!! See the animation (contains large amounts of "jitter" due to the different camera viewpoint in each frame) and adjacent images for object identification. Read more...

Yesterday: After Friday's struggle differentiating the maximal clique algorithm from my approximation approach, I determined that there doesn't appear to be much difference unless its application-specific. Have started implementing/adapting the maximal clique algorithm to speed up based on the conditions for 3D rigid object detection. Cursory results seem promising, though a wider selection of test cases needs to be explored.

Yesterday was spent getting Perera's implementation up and running (as it tends to). C-based implementation for maximal clique detection seems fast, but the datasets are small (and correspondences pre-determined). Requires further evaluation. Emailed Faramarz about discussion/direction.

Today: Start comparing and contrasting the maximal clique approach: re-arrange personal datasets for input into Perera's and similarly Perera's datasets into my own. Compare/time differences in maximal clique approaches/short-circuiting. Discuss/talk out with Faramarz how the approximate clique might be used for speedup. Read more...

Yesterday: (last couple days) Continuing with paper writing - very slow progress. Started writing up the methodology. May need to re-evaluate how I'm using the approximate maximal cliques, and how best to justify it.

Today: Continue re-evaluating exactly how the approximate maximal cliques is being used, and justification for why its worthwhile in this situation. Write up how the distance matrix works, what it achieves, etc. Include the notes from Faramarz. Read more...

Yesterday: (last couple days) CPSC453 marking for the raytracer. Couple more paragraphs in background and onto methodology.

Today: Finished marking. Read more...

Yesterday: (last couple days) TA demos. Doctor's appointment. Couple paragraphs in background covering graph representation and maximal cliques.

Today: Wrapping up background work on static paper including mathematical descriptions and strong citation sources. Starting algorithmic representation. If writing motivation dwindles, work on Kinect incorporation/consider how to represent distance graph probabilistically. Read more...

Yesterday: (last couple months) Still not dead, but I temporarily lost sight of a scrum being short rather than long-winded and a giant todo list.

• TA responsibilities, including marking
• Identified areas for parallelization in the system, including ability to dynamically adjust for grouping analysis based on time constraints
• Generated synthetic good examples, as well as a presentation of the work.
• Connected with Bio researcher for animal input/ideas (+ volunteering for a stationary wildlife camera team)

Today: Answered a number of TA questions. Wrapping up background work on static paper. Read more...

Yesterday: (last month) Not dead.

• Been looking into how to save vertices based on what is visible - still struggling with that, though I have a few ideas, and may just sit down one evening with a movie and workout the poor approach of doing this.
• TA responsibilities, including marking
• Struggle, but seems to have come out alright the other end, with articulating the problem to the supervisor. Summary is that we're working with a variant of the maximal clique problem. The approach I'm taking aims to simplify the solving so its not as extensive. Regardless of which approach taken to identify the cliques, we still need to do some post-processing on the resulting cases: where distances were (briefly? sometimes continuously) maintained and there is overlap, or where a common edge is shared.

For example, for 49 polyhedrons bouncing on a table, we have a number of points that should be distinct, but exhibit overlap with other vertices, and therefore have unique clusters (myPhysicsReduced). Unclear if this is due to the large number of objects, or because within the example, many objects bounce with a common pattern.

Today: CG supervisor meeting, including two core take-away implementations to work on: 1. what happens when we define our adjacency graph based on consistency in edge length history (see image above); and 2. what cases need to be considered for object labeling (ie. pattern overlap recognition, ala clique-finding). Given the implementation for 1, the current bottleneck may be related to loading the data (eg. +10,000 points). Read more...

Yesterday: Starting to develop a Blender script for outputing vertices to file. This way non-visible vertices will be flagged as null, but still present/in-order.

• have low expectations for how much is created for the resulting objects, based on the current example scenes
• ... therefore consider building a synthtetic example where it moves in front of the camera
• have vertex visibility based on camera? current display view? (which means I need to set a camera within the scenes
• re-work scenes so that the camera is animated around the objects for increased object observation

For the matlab code:

• adapt the import sections to accept these text files, and select the subset of vertices that are visible for a frame to work with
• this means that when selecting the current visible vertices, they may not have been visible in prior frames - adapt the code to accept that there is no history
• for vertices that have previously been assigned an object but are not currently visible, update their transformations based on how the object has changed. options include:
  • object remained unchanged - apply the same transformation for where we expect that vertex to be
  • object has split - flag the vertex has being invalid/inconsistent/unknown... future dealings with this vertex should observe which object it was previously associated with, and how that object was divided over time (time-graph representation of object construction/destruction and the observed transformation for each frame?)
• be able to select and show the developing state of a given object - ideally would be able to rewind and see how it has changed; or alternatively rewind and re-incorporate the more expansive object into transformation stream

Today: TA preparations and tutorials. Identify what needs to be done for potential position application (resume, Gallery? updating). Continue building the codebase for incremental updater. Flag for performance and accuracy (high object count) improvement the different sections within matlab, and be able to justify why the datastructure thus far is of less importance for tomorrow's supervisor meeting. Read more...

Yesterday: (General) Been busy with getting tutorials up and running for CPSC 453 and after assistance from a colleague, have been developing object segmentation from random points. At this stage, we have successful segmentation of: double pendulum, five polyhedra on table, two spheres on table and engine parts. I can also visualize the respective parts separately once they've been generated.

Competed and submitted application for Graduate Scholarship. Finally met with supervisors last week, and discussed possible improvement directions. While our discussion suggested exploring data structures for improving limitations to vertices (eg. using sparse matrix representations), some time hang-ups (engine takes 30 minutes, 49 spheres on table takes a large chunk of a day and is erroneous) might be achieved by better algorithm use. Future discussions are anticipated to explore effective use of history for fast segmentation.

(Friday) Replaced a number of for-loops with (faster) inline matlab commands, and stepped through why there might be issues with object segmentation for scenes with a high number of objects.

Today: Predominantly preparing for tutorials today. Will continue exploring issues with ineffective object segmentation. Tomorrow will start generating scenes with obscured vertices, and be able to read in such scenes. Read more...

Yesterday: Started application for Graduate Scholarship, coordinating references. Follow up and resolving payment of Pyxis 2014 workshop. Coordinating meetings with supervisors, and biological modeller Mike Terry at UWaterloo.

Stepped through codebase where too many objects are being identified. It ends up being multiple (4 or more) vertices mapping to the same adjacency pattern, which has strong overlap with what should be a cohesive object. Need to determine how to differentiate.

Today: Look into bugs from yesterday. Follow up on meeting scheduling. Have consistent colouring/labelling across frames. Start of term TA meeting. Read more...