> I stopped doing CV research because I saw the impact my work was having. I loved the work but the military applications and privacy concerns eventually became impossible to ignore.
And its conclusion is gold:
But maybe a better question is: “What are we going to do with these detectors now that we have them?” A lot of the people doing this research are at Google and Facebook.I guess at least we know the technology is in good hands and definitely won’t be used to harvest your personal infor-mation and sell it to.... wait, you’re saying that’s exactly what it will be used for?? Oh. Well the other people heavily funding vision research are the military and they’ve never done anything horrible like killing lots of people with new technology oh wait.....  I have a lot of hope that most of the people using com-puter vision are just doing happy, good stuff with it, like counting the number of zebras in a national park, or tracking their cat as it wanders around their house. But computer vision is already being put to questionable use and as researchers we have a responsibility to at least consider the harm our work might be doing and think of ways to mitigate it. We owe the world that much.In closing, do not @ me. (Because I finally quit Twitter).
 The author is funded by the Office of Naval Research and Google.
"pjreddie 30 points·2 months ago· edited 2 months ago
I've never been in a car that was using my tech to avoid killing people but I have had a 3 star general rave about how my work was being deployed in war zones and how army research groups love my software.
Edit: to say that i'm not arguing against CV research as a whole, i'm just saying i don't want to do it anymore because of the impact i saw my work having"
There's not even any need to go as far as assume evil intent on the users of the software, just plain recklessness can easily cause people to die.
Also consider that object detection has a lot of impact in positive ways as well. It doesn't seem so black and white.
I'm a researcher working on a system for monitoring offshore kelp farms for renewable aquaculture with an autonomous underwater vehicle. I can't use GPS underwater, Dopper velocity logs and sensitive inertial navigation systems are either too expensive and/or export-controlled, and doing manual filtering of visual data is tricky and inconsistent. Adding good object detection for kelp to helps to make that kind of system much more reliable, which can help provide useful metrics to groups working on creating new biofuel sources and marine ecosystem monitoring. I think that's valuable work, and it's enabled by YOLO.
Joe Redmon has paid attention to where and how his work is being used, and I respect him for disengaging with something that he finds not to line up with his values. But it's worth pointing out that there are people (not just me) who are using that work in ways that might be worthwhile.
Your work seems worthy of an AMA if you ever have the time.
For example, the robotics end is currently a human-piloted ROV with data getting obtained in post-processing on dozens of fronds, rather than an AUV doing real-time inference (hopefully to be solved by YOLO) on thousands. However, I recently heard about some tentative plans for a more in-depth pilot-scale effort in Spring/Summer 2021, by which time we'll hopefully have some more interesting results to talk about!
I'm curious if this is something that some sort of modified license could help resolve. Do I have the option to license my software so that it can't be used in war?
"The Software shall be used for Good, not Evil" - https://en.wikipedia.org/wiki/Douglas_Crockford#%22Good,_not...
Licenses are not a viable solution to this sort of ethical quandry.
Or do you sigh only because in this instance you agree with the original author, but you don't actually think it should be a rule that others are constrained by the moral thoughts of their predecessors?
I know of a promising treatment for types of hypothyroidism, but the original discoverer doesn't want to continue work on it because she doesn't agree with animal testing.
If my moral calculus says that the quality of life of millions of humans is more important than the quality of lives of thousands of rats and dogs, am I not allowed to pick up where she left off?
I often use my 'Stabby the Robot' diatribe to convey this to people. Imagine a robot that moves too fast to evade, that doesn't rely on expendable, limited ammunition(i.e. a blade or sharp surface), and can locate the jugular vein of a human. You could make one now out of a commodity drone, but it wouldn't last very long because of the power problem. AI would obviously make it more dangerous, but it isn't necessary.
We've already placed very dangerous tools into the hands of humans. Maybe it helps AlexeyAB sleep at night to stop working on YOLO, but his idle hands are not holding back the future.
Oh hey I remember that write-up on qntm!
There's a reason the US military has been playing with drone clouds for years.
I rogue state, or even a non-state actor, can leverage cheap drones(assuming, of course, a solution to the power problem) to obtain a weapon of mass destruction. Drones are fundamentally different from bullets.
and I suspect there are dozens of other similar projects out there.
Mitch: You did?
Chris: Yeah, and he used to be the number one stud around here in the 70’s. (whispers) Smarter than you and me put together.
Mitch: So what happened? Did he crack?
Chris: Yes, Mitch. He cracked, severely.
Chris: He loved his work.
Mitch: Well what’s wrong with that.
Chris: There’s nothing wrong with that, but that’s all he did. He loved solving problems, he loved coming up with the answers. But, he thought that the answers were the answer for everything. Wrong. All Science no Philosophy. So then one day someone tells him that the stuff he’s making was killing people.
Mitch: So what’s your point? Are you saying I’m going to end up in a steam tunnel?
Method | Top 1 | Top 5
No-op | 78% | 94%
Swish | 64.5% | 86%
Mish | 79% | 94.5%
Swish is the only value that decreases performance (and by a huge magnitude) but a very related methodology improves performance hummm...
Article on YoloV4