The Camera Is the Lidar
medium.com
The Camera Is the Lidar
1–10 of 77 posts
Re: The Camera Is the Lidar
#2Is it just me, or are we seeing yet another impressive leap in Computer Vision that's soon going to be hyped as an incremental step into Skynet?
Re: The Camera Is the Lidar
#3This is so brilliant, I'm considering completely switching the type of software/hardware work I've been doing for the last decade.
Re: The Camera Is the Lidar
#4The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.
Re: The Camera Is the Lidar
#5Is it just me, or are we seeing yet another impressive leap in Computer Vision that's soon going to be hyped as an incremental step into Skynet?
Unless you think Skynet is impossible, all impressive leaps are an incremental step towards Skynet.
Re: The Camera Is the Lidar
#6The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.
The status quo is neural nets already are allowed in safety critical systems - humans.
And with the amount of drivers snapchatting behind the wheel, I'd rather take my chances with a self-driving car.
Re: The Camera Is the Lidar
#7Something i don't understand : i've seen those kind of videos (i think it was nvidia or maybe waymo) where a signal (either camera or lidar or both) is processed in real-time and displays boxes around cars, street lights, etc. It seemed to be that detouring real-life objects from sensors in real-time has been working for a long time now.
What does this bring that's new ?
Re: The Camera Is the Lidar
#8The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.
The status quo is neural nets already are allowed in safety critical systems - humans. And with the amount of drivers snapchatting behind the wheel, I'd rather take my chances with a self-driving car.
> neural nets already are allowed in safety critical systems - humans.
I’d really like to see a demonstration that human behavior is just neural nets. Sadly, I think that’s still an open question.
Re: The Camera Is the Lidar
#9The problem is that deep learning should not be allowed in safety critical systems, because (1) the accuracy is always less than 100% even in known test situations, and (2) we don't know how it works and under what conditions it breaks down.
Every self driving car company is using neural networks. If neural nets are going to be used anyway, I'd rather have centimeter level range information per pixel than not.
By the way, although the article focuses on deep learning, there are many applications that don't involve deep learning. For example, although you can run the deep neural network based SuperPoint on the intensity data, you can also run any classical feature extraction algorithm such as SIFT, SURF, ORB, BRISK, FAST, AGAST, etc. Doing so provides an elegant solution to the problem of localizing in a geometrically sparse but visually rich environment, such as a smooth but well-illuminated tunnel.
Re: The Camera Is the Lidar
#10Something i don't understand : i've seen those kind of videos (i think it was nvidia or maybe waymo) where a signal (either camera or lidar or both) is processed in real-time and displays boxes around cars, street lights, etc. It seemed to be that detouring real-life objects from sensors in real-time has been working for a long time now. What does this bring that's new ?
Estimating depth and 3D bounding boxes using cameras only, while possible, is much less accurate than using actual range data from a lidar.
On the other hand if you fuse lidar and camera data such as with Waymo and others, there may be issues with the sensors being out of sync (as they run at different framerates, and the lidar continually spins) or physically offset (leading to parallax issues). Dealing with such issues is very difficult. Having a single sensor output both accurate range information and camera data makes it much nicer to work with.