Undergraduate Thesis Topics

rexo12

I'm trying to get a Computer Science honours project (essentially a year-long undergrad research thesis) off the ground, looking at something in computer graphics and realtime rendering. I have a project idea from my work that I have discussed with a supervisor, although it looks like it'll be a deep learning problem which I'm not super interested in.

As such, I'm looking to get a bit of a survey of some other open questions in real-time graphics that would be good to look at, and maybe some resources to do further research on. I am interested in material definitions (BRDFs and such) and acceleration structures like distance fields.

Thanks!

Eric Chadwick

That’s exciting, lots of options for research and development in real-time rendering. So many in fact that it may be difficult to recommend anything specific except what my own biases would suggest

I’m really intrigued by augmented reality, since we use it at my work, and there’s a lot of hardware development happening in this area, like the recent Apple device, the Magic Leap system, etc.

I would suggest searching for AR videos on Twitter and LinkedIn, lots of people posting short TikTok like videos of neat r&d experiments.

A couple examples 
https://twitter.com/Alientrap/status/1574551281620111361

https://twitter.com/pushmatrix/status/1328709526426701829?s=21

https://www.linkedin.com/posts/endritrestelica_technology-medical-healthcare-activity-7078405162188333056-wTjf

poopipe

If someone came up with a performant way to render foliage and/or hair as a volumetric effect* that allowed an artist to author content in a controllable and reasonable way they'd probably win some sort of prize (I'd also try quite hard to hire them) . 
just sayin.. 


*I suspect there's a reason it hasn't been done

Eric Chadwick

Oooh, maybe NeRFs could be bent to do this.
https://www.matthewtancik.com/nerf
https://cseweb.ucsd.edu//~viscomp/projects/LF/papers/ECCV20/nerf/website_renders/lumpyroots_200k_rgb.mp4
https://cseweb.ucsd.edu//~viscomp/projects/LF/papers/ECCV20/nerf/website_renders/fullmagnolia_200k_rgb.mp4

rexo12

Eric Chadwick said:

 So many in fact that it may be difficult to recommend anything specific except what my own biases would suggest

That's precisely what I want to hear about!

AR is a big part of the stuff I do at work, a lot of the current problems seem to be in calibration & synchronisation (and a lot of the solutions, given the fuzzy input space it needs to work with, are probably deep learning related).

A number of the papers published at SIGGRAPH over the last few years seem to be related to 'Differentiable Rendering', which is kind of magical as a concept although hardly realtime. It's a bit frustrating that so much of computer science is interested in neural or deep learning techniques at the moment, since I find them not that interesting to study. even the topic linked above typically employs some form of gradient descent.

Being a computer science project I would have to approach whatever problem I do from a theoretical angle as opposed to an engineering one - designing some new algorithm or data structure or improving on an existing one somehow.

@poopipe would you be able to elaborate on what you meant by a "Volumetric Effect"? As in literally modelling these objects as volumes rather than cards/strands - something like this?

poopipe

umm,  sort of.  
if you ignore the neural bullshit it's doing the same thing your standard offline hair rendering does. You define a small number of guide splines and it fills in the gaps - in that sense yes,  the artist defines the volume and the machine does the rest.

The harder problem is transferring that principle to foliage because leaves aren't just splines.

The reason I'm actually interested in this is that the foliage and hair models generally used for games at present do not scale well.
It is very difficult and labour intensive (i.e costly) to produce a card based tree or hairstyle that smoothly transitions from one level of detail to another without noticeable sillhouette and lighting (surface normal) changes.   

The theory is that if the model is represented as a volume then you are better placed to eliminate both these issues - you simply draw less strands/leaves as the object becomes smaller on screen and can reference the original volume for lighting information
This theory is borne out to an extent by the pseudo-volumetric fin and shell technique used in several games - it isn't well suited to certain hair styles/types eg. ponytails or curly hair but it is scalable without artefacting and very easy to author for. 
Planet Zoo is a good example of this technique being pushed to it's limits and I believe the recent Ratchet & Klank game uses a similar method(don't quote me on that one). 
Nothing like that exists for for high detail foliage to my knowledge

rexo12

@poopipe, I see, thank you for the writeup! That's quite an interesting problem.