To me, this is whether you have good people around or not. PhD usually comes with good advisers who help you think in different ways than you have done before. However, such good advisers can be also at a company + you work on practical an modern problems (usually). For that you don't need PhD. In any case, once you ask whether you should do PhD or not it feels like you don't want to be left out. In that case, just do it.
Consider for example a picture of a street and a rendered car overlaying it. It would be interesting to see how well the technology bakes in the car so that it matches the environment. This would go in direction of realistic rendering.
4. This statistics must be conditioned. Not all drivers are the same. Some are careful, some are not. So in fact, the careful drivers will make their life miserable by using autopilot.
When the self driving revolution started, conservative people claimed that one accident would stop the whole business. Well there are many accidents already and people are still willing to risk. One reason may be our tendency to trust something which we do not understand. We feel that if the feature is already in the car, it must be safe. Well it is not. And even worse, it is driven more by marketing and money than by safety regulations.
I am very skeptical about pretraining which seems to be the key point of Nanonets. Sure, it will help to work work better than initialization from random weights, but, you will always do better if you collect more data for your problem. This may be fine for problems which do not need optimal classification and fast performance, but I am struggling to see any use case for that.
Can you link to the images you have tried? My tests were rather disappointing. Yes, I got some 3D point cloud from the image, but nothing that I would consider useful outside computer vision community for now.
You are right, there is good application for this tool in education or people who want to quickly test something new. Shadertoy is great and I like it. It is at the right level of abstraction that I need for my current CG knowledge. I can imagine that there are many people who find Live CV very useful and then move downwards.
Prototyping YOLO or the other DNN means just run detector and see the boxes. Of course training these systems is clearly out of reach for this kind of abstraction technology, not to mention modify and develop something new.
This can never work in practice. First, it seems super simple to do CV in the video, but only because the author already knows what to type there. It will never be so smooth in real life. Next, if you accept that you have to think what you are doing, why not to fall back to OpenCV in combination with ImageWatch visualizer. The good news is that you save yourself one layer of abstraction which hides important details from you, not to mention learning yet another API.
I like the separation of glasses from compute power. The design itself seems cool, a little bit inspired by Ghost in the Shell. But imagine putting this on average dude, the coolness is gone ;)
Agreed. The network is trained in a sense "online" on the target image and directly applied to it. I don't see much impact of this work outside image denouncing or tasks already presented in the paper.
I would like to see how the system generalizes on different videos. The motion pattern seems to be quite general ... this can be used to analyze crashes during racing events for instance (http://www.tldvision.com/tld3.html).
Data scientist scores top in the chart. Funny to find later on that the author calls her self data scientist as well. To me this is little off putting. It is like a doctor claiming that doctors are the most healthy people. Which might be true, but strange, don't you feel? ;)
Already mentioned in the comments. But still. Looks nice, but it is hard to get any insight from it. For instance, I zeroed out all weights and outputs, then slightly increased one weight. First there seems to be some non-zero bias on the minimal weight, second there is some mon-linearity which goes against intuition. It would be nice to be more clear about what is going on and then it can be more useful not only nice.
Sure, we are able to work/communicate remotely. But the problem is that even remote work on super-computer is limiting and reduces the iteration cycle. I prefer to run small and fast experiments on local machine, fine-tune for significant time and after this send them to big cluster. Once in this stage, observing the progress from coffee shop is more like a "gradient addiction" for me. Why? On local machine I get much faster iterations and find errors quickly. In Elon's words, the "bandwidth" between me and machine is higher. The merge started, but there is still a lot to be discovered in order to broaden the information flow.