I had two questions regarding the road segmentation project.
(Scientific novelty): The instructions for the projects say that our method should be "creative" and "scientifically novel". I have been scratching my brain for some time now, and any time I think of a novel idea I find a paper that already does it anyway. So my question is: how important is it that the method we try is new? If for example we implemented a fancy method that is already published in some paper (citing it, of course) would that penalize our submission in the "novelty" grading criterion?
(Quality vs quantity): Is it more important to create one single model that performs very well, or do you want to see us implement a lot of methods that do "okay" ?
Hey ! I just answered your email but here's what I anwered:
For your 1rst question, we don't expect you to come up with totally new ways to solve the task, especially for the challenges as they are already well known problems. You can use well known models and try to tweak them a bit with different losses, regularization, or whatever you have in mind! (But if you manage to find a totally new way to solve the task, sure go ahead)
For your second question, the best way to go, in my opinion, is to pick a baseline and then try new things to see if it improves the performance.
As a baseline you could use something from the "tutorials" of the project for example (regression, CNN). Then, it is up to you, you could implement very different models, or pick only one that you try to optimize in different ways. But as the SOTA models can be really computationally expensive, I would not advise you to implement more than 2 of them (unless you've got a farm of gpus that is).
Also one thing that is quite good for the report, we should see that you built your way up to get your final preformance, hence the baseline :)
Questions about road segmentation project
Hello,
I had two questions regarding the road segmentation project.
(Scientific novelty): The instructions for the projects say that our method should be "creative" and "scientifically novel". I have been scratching my brain for some time now, and any time I think of a novel idea I find a paper that already does it anyway. So my question is: how important is it that the method we try is new? If for example we implemented a fancy method that is already published in some paper (citing it, of course) would that penalize our submission in the "novelty" grading criterion?
(Quality vs quantity): Is it more important to create one single model that performs very well, or do you want to see us implement a lot of methods that do "okay" ?
Thank you for your time.
Hey ! I just answered your email but here's what I anwered:
For your 1rst question, we don't expect you to come up with totally new ways to solve the task, especially for the challenges as they are already well known problems. You can use well known models and try to tweak them a bit with different losses, regularization, or whatever you have in mind! (But if you manage to find a totally new way to solve the task, sure go ahead)
For your second question, the best way to go, in my opinion, is to pick a baseline and then try new things to see if it improves the performance.
As a baseline you could use something from the "tutorials" of the project for example (regression, CNN). Then, it is up to you, you could implement very different models, or pick only one that you try to optimize in different ways. But as the SOTA models can be really computationally expensive, I would not advise you to implement more than 2 of them (unless you've got a farm of gpus that is).
Also one thing that is quite good for the report, we should see that you built your way up to get your final preformance, hence the baseline :)
Add comment