Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Facebook open-sources Detectron

research.fb.com

Post reply on HN
Print view View original

41–50 of 188 posts

Re: Facebook open-sources Detectron

#41

Post by m_ke »

deepGem wrote: post #38
m_ke wrote: post #37

Earlier quoted context omitted.

You could try tensorflow object_detection api with tensorflow lite https://github.com/tensorflow/models/tree/master/research/ob... google also recently put up their mobilenet v2 paper which handles segmentation https://arxiv.org/abs/1801.04381

+1 for the Google object detection API. The trained model is quite huge though. 200 MB based on Resnet faster R-CNN. There are creative ways of chunking this model to keep it small.

I think they have a mobilenet SSD model as well.

Re: Facebook open-sources Detectron

#42

Post by m_ke »

yters wrote: post #40

Is there a class of math problem humans can solve but computers cannot? Then we could just use these problems as a guaranteed test instead of the current CAPTCHA arms race.

There is no arms race. 99.9% of the time Google knows if you're a robot based on your browser state. They make you label the images because it's a free way to get training data.

Re: Facebook open-sources Detectron

#43

Post by tvmalsv »

itsa2 wrote: post #22
m_ke wrote: post #17

Earlier quoted context omitted.

Not that much worse than what you see on imagenet. Most large companies have internal datasets with >100 million images. https://arxiv.org/abs/1610.02357

In case you didn't realize, the guy you're talking to is the CTO of what some may call a "large company."

And I enjoyed the conversation between two people that obviously know the field pretty well.

Re: Facebook open-sources Detectron

#44

Post by ftoo »

m_ke wrote: post #42
yters wrote: post #40

Is there a class of math problem humans can solve but computers cannot? Then we could just use these problems as a guaranteed test instead of the current CAPTCHA arms race.

There is no arms race. 99.9% of the time Google knows if you're a robot based on your browser state. They make you label the images because it's a free way to get training data.

How does that work? Could we create an open source lib to replace Recaptcha?

Re: Facebook open-sources Detectron

#46

Post by pilooch »

franciscop wrote: post #36

Does anyone know an alternative that works on RaspberryPi? This states: "Detectron operators currently do not have CPU implementation; a GPU system is required." Even low FPS (3-5) would be acceptable.

See https://github.com/jolibrain/dd_performances you can reach 1fps on image classification and objects detection tasks on rpie3 with custom caffe and models.

Re: Facebook open-sources Detectron

#47

Post by stmw »

chewxy wrote: post #28
stmw wrote: post #11

This is great! I do wish this were written in something other than Python. What is the carbon footprint of all this computer vision, compute-intensive code still being run billions of times a day in Python? Someone should calculate...

Golang alternative being developed (by me and a bunch of others): https://gorgonia.org/gorgonia

This is very cool, I hadn't heard of it!

Re: Facebook open-sources Detectron

#48

Post by stmw »

linkmotif wrote: post #31
stmw wrote: post #24

Earlier quoted context omitted.

It is funny to see this comment get "-4" already... What's so offensive? After all, Facebook has rocksdb in C++, percona in java, and a PHP->C++ compiler, so they clearly have both the belief and the skill in moving away from interpreted programming languages for performance-sensitive code.

For some reason, people are offended by gross misunderstandings. This framework is in Python, but it’s a Python binding that sets up code that runs natively (not even sure the details myself; others are writing CUDA). TensorFlow is the same way. It’s in Python, but the computations are not in Python. As you point out, that wouldn’t make sense.

Having written Python -> C bindings before, I am quite aware of how this works... It still has considerable overheads.

Re: Facebook open-sources Detectron

#49

Post by stmw »

inlined wrote: post #20
stmw wrote: post #11

This is great! I do wish this were written in something other than Python. What is the carbon footprint of all this computer vision, compute-intensive code still being run billions of times a day in Python? Someone should calculate...

What was the carbon footprint of the turk machines the Python can replace?

are you seriously comparing inefficient python code to humans?

Re: Facebook open-sources Detectron

#50

Post by visarga »

m_ke wrote: post #29
schrep wrote: post #21

Earlier quoted context omitted.

That’s impressive work. Still don’t think we have reached human level for all the categories of things we see in images. But you are correct that my comment about 1k categories is not true for many production systems.

Definitely not close to having things work for all categories. As you scale up to more categories ambiguity and specificity becomes an issue. Clarifai has a nice demo of their model which has >10K classes, https://clarifai.com/demo , the top predictions are usually correct but not always the most relevant. I only linked to the xception paper because it mentions JFT. It's not state of the art for large scale recogniti…

It's not just a matter of detecting objects and locating them. The deeper computer vision problem is to identify object attributes, relations between objects and actions in video. It's much harder to do that because many relations appear in very diverse situations, with objects of different categories, so it's hard to have 1000's of examples for each class of relation.

For example, humans can identify a monkey riding a Segway on the airport runway, but there probably is no such thing in the training set, even if it is quite large. The neural net might not know if that constitutes a "riding" action because it has never seen such a combination. Maybe the monkey is jumping over the thing and the picture shows it in proximity to it, not riding it - a human would know that a slight gap means there is no riding taking place.

Then, the even harder problem is to predict the consequences of actions on objects and just to physically simulate the scene. Such knowledge is useful in robot action planning. Beyond computer vision, there is also a need to create a "mental simulator" that has theory of mind and can simulate other agents (what humans intend), and we need simulators, both physical and mental to create the next level of AI.

Post reply on HN