mathgaron · HackerTrans

I helped build a similar setup a "long" time ago to build a sky HDR database. We had over 40k images across hundreds of days. The longer exposures definitely did not help (even with a ND filter) and by the end the camera had some broken pixels along the sun trail. Nothing noticeable visually, but likely not good for computer vision purpose. If we were to redo this it would be definitely a valid concern to address.

mathgaron

7 anni fadiscuss

I don't think that a bad employee is someone that lacks experience. The article is not so clear on that point but I don't think it is what the author meant... That said, in my opinion, a bad employee is someone that gains nothing from experience and that is not going forward, and this can easily be detected after some time.

mathgaron

7 anni fadiscuss

Those cameras were developed by an Israeli startup that Intel bought a few years ago (Omek Interactive). Making these sensor low power could enable devices to have 3D reconstruction and tracking. This opens the door for more robust AR/VR applications and cheap devices for robotics. Not only that Intel is in a good position to sell/produce those sensors, but they probably sense that it could be a good market to conquer.

mathgaron

7 anni fadiscuss

Nice project! Your 3D printed case for the stereo pair removes a lot of potentially useful information for the feature extraction algorithm. If it can help, I have an old project where we mounted a realsense (R300 if I remember well) on Hololens.

You can inspire from our 3D printed mount, the CAD is available here : http://vision.gel.ulaval.ca/~jflalonde/projects/hololens3d/i...

mathgaron

7 anni fadiscuss

Great example of Baader–Meinhof effect. When you recently hear something and you remark it everywhere. I can assure you that I also read this post and did not talk about crows/raven recently!

mathgaron

7 anni fadiscuss

I guess it is no secret that what helps is a schedule.

I do it with accordance with a to-do list: setup realistic deadlines for all projects/side-projects and build a schedule based on them (The deadlines are the only "tight" stuff in my schedule, the rest is flexible). Your work/life balance has to be part of whatever you plan! (e.g. I usually try to have ~2 weekends in the month for other activities and ~2 weekends for projects).

Also it helps me during my free weekend/night to simply keep away from my laptop, this way I don't fall in the trap of wanting to work 1 hour and end up working the whole day.

Sadly your problem is quite common. I know too many people that can't handle their work/life balance...

mathgaron

7 anni fadiscuss

You can still switch your focus with these because you still see the real world. However it is true that the objects are rendered ("OpenGL-like") the same way no matter where the object is located in the scene, so while it gives you the feeling that you can focus your eyes on the object (because of their 3D position and the tracking), the objects always appear "clear" which can be annoying. That said, using them 8 hours must be hard : the virtual objects emits unreal lighting (different and stronger than the light reflected by real surfaces)

mathgaron

7 anni fadiscuss

I talk from my personal experience with V1 and other devices using waveguides display, I felt like it is not the best type of display to do "heavy work" such as reading/coding. It is quite annoying to have everything ghostly, even if the tracking is stable and the resolution is good.

mathgaron

8 anni fadiscuss

Are you talking about a depth channel on an image plane? It also largely depend on the problem but here is a few tricks that helped me for my problems (object tracking).

- Generating synthetic data is powerful if you have the depth modality as it is easy to render. Also the real/synthetic domain gap is narrow compared to RGB. I consider it as data augmentation: you usually do many renders from a single 3D model.

- If you can somehow normalize the offset (e.g. compute normals) that can help. In my case I could offset the center of the object as 0 depth and it greatly help the network to converge.

- Classic augmentations like gaussian noise, gaussian blur and also downsampling the depth helps (apply these randomly).

As for tooling, I just use numpy/pytorch for most operations and OpenGL for renders.

mathgaron

9 anni fadiscuss

The IR camera is near-infrared, it measures more or less light that your eyes can't see, and no heat frequencies. The IR camera is used to recover 3D information: https://en.wikipedia.org/wiki/Structured-light_3D_scanner

comments