Sniper – An efficient multi-scale object detection algorithm | Hacker News Reader

Sniper – An efficient multi-scale object detection algorithm

github.com

156 pointskumaranvpl33 comments

Open articleView on HN

that's uh... quite a name...

I understand it was probably meant tongue in cheek. Still, very unfortunate naming.

Just be thankful it was this project and not the upcoming face detector mentioned at the bottom of the readme.

Perhaps he'll call that "Crosshair".

Well his other project is called SSH

Ha. I thought you meant he was the creator of THE "SSH". Nope. It's just another machine learning library.

I guess it's a reference to how the algorithm zooms in on a smaller region of the image before picking out an object.

And killing it

Then how about "Zoomer" instead?

It's a really really bad name for an object detection system

Why? Let's be honest, the main use for most of this new ai tech is going to be military.

Why?

Cars, household-robots, etc. all need to get around as well.

Sometimes you just gotta aim at the head.

Re: "Let's be honest, the main use...going to be military."

While perhaps true, usually one finds a friendlier name regardless, in order to hide or reduce the perception of association with war. It's why we have a "Dept. of Defense" instead of a "Dept. of Offense" or "Dept. of War". I'm not necessarily condoning such a practice, merely making an observation of product and project naming practices.

We used to have a Department or War

You still do, its just given itself a more palatable name..

I see that Uber was using this in Az back in March...

It's not named babyseekingbrutalkiller9000, meh.

Darnit, you guessed my password ;-)

It's an acronym: Scale Normalization for Image Pyramids with Efficient Resampling.

So up next is "Bi-Linear Object Oriented Detection"?

Unfortunate how CUDA has such a stranglehold on scientific work

Why is that bad?

When all you have is a hammer...

I'm guessing Nvidia hardware lockin

Competition is welcome, deep stack knowledge from driver to model building is hard to come by (it's more than just matrix multiplies).

To compete with CUDA you need high performance drivers and people that can write them, and also be able to anticipate and keep up with the future needs of ML as well as hardware development.

The current title refers to the algorithm as "Sniper", but the name appears to actually be "SNIPER". According to the paper [1] linked in the repo, that stands for "Scale Normalization for Image Pyramids with Efficient Resampling."

[1]: https://arxiv.org/abs/1805.09300

Tldr would this work on a pi3?

That'd be interesting to know. Could you try it and report back?

I'm no expert on rpi or cuda, but is it actually possible to run cuda code?

No.

Given that it gets around 5 FPS on a V100 gpu I doubt it can run on a raspberry pi.

Reply on news.ycombinator.com