yeldarb

I’m not sure if this is real, but this video of a stable diffusion powered video editor was posted on Reddit a couple weeks ago and looks insane: https://www.reddit.com/r/MachineLearning/comments/wmypmh/a_d...

Or flamethrower https://hackaday.com/2021/09/11/dont-sleep-on-the-lawn-there...

We summarized the high level improvements here: https://blog.roboflow.com/yolov7-breakdown/

For object detection, we made a way to do it from scratch in a couple of hours (for free if your project is open source[1]): https://docs.roboflow.com/inference/web-browser

[1] https://blog.roboflow.com/computer-vision-datasets-and-apis/

Yes, this is called "instance segementation". There's a YOLO-based instance segmentation model called YOLACT.

Hey all, excited to share that we've opened the floodgates on Roboflow Universe and now host the world's largest collection of open source computer vision datasets and pre-trained models, shared by our community.

Last year we made our core product (focused on building custom computer vision projects) free for hackers and hobbyists working on open sourced projects and tens of thousands of developers jumped onboard. Now there's a starting point for basically any computer vision use-case you can imagine.

For example, yesterday I tagged a bunch of Robotics datasets[1] users have shared. We've also tagged other collections[2], but there are tens of thousands more projects that you can find using the search (pro-tip, add "trained model" to your search to filter down to only projects that have a pre-trained model included -- you can use these via API or deploy them to run on edge devices).

Super excited to see what developers will be able to build now that they don't have to start from scratch (and may not even need to collect and label data) to add CV to their projects.

[1] https://universe.roboflow.com/browse/robotics

[2] https://universe.roboflow.com/browse

Can't be this; Google Colab gives out tons of free GPU usage.

Any interest in over engineering it a bit more? We did something similar to spook rabbits[1] in a garden using a computer vision model to detect them (vs your masked motion sensor). Looks like lots of users have shared pigeon detection datasets[2] that could make a good starting point.

[1] https://blog.roboflow.com/rabbit-deterrence-system/

[2] https://universe.roboflow.com/search?q=Pigeon&t=subject

Only if you purchased the optional "earmuff" peripheral & have installed the latest MUTA drivers.

Is there a compatibility layer or migration guide for people looking to switch?

The In Sente YouTube channel has a good series of videos that cover each of the chapters from this book: https://youtu.be/6YvKJN2qso4

Roboflow | Engineering, Community | REMOTE (USA)

Roboflow (roboflow.com, YC S20) builds tools that help developers give their software the sense of sight. You can train and deploy a computer vision model in an afternoon without having to become a machine learning expert.

Over 100,000 developers (including employees of over half of the Fortune 100) have used Roboflow to build everything from self driving cars, to pill counting mobile apps, to cat laser turrets[1].

We recently raised our $20M Series A[2] and David Sacks (Paypal, Yammer) joined our board. And we were just listed as a top YC company[3].

Our business has grown much more rapidly than our company & we are expanding our engineering team to keep up. We're looking primarily for Full Stack Engineers to work on our core platform (consumed via web browser) and code tentacles (consumed via code). There are tons of interesting problems to solve & you will have a huge amount of impact and autonomy.

We're dedicated to building a world class remote-first culture[4] and becoming a magnet for exceptional talent. Check out our available listings[5] (5 more engineering roles being posted this week), try out the product[6], and get in touch!

[1] https://www.youtube.com/watch?v=AVp1tndzV34

[2] https://blog.roboflow.com/series-a/

[3] https://www.ycombinator.com/topcompanies

[4] https://blog.roboflow.com/remote-not-distant/

[5] https://roboflow.com/careers

[6] https://app.roboflow.com

Almost certainly.

Just like you don't need to become an expert in cryptography to add auth to or accept payments in your app, you won't need to be an expert in machine learning to make intelligent content recommendations or add contextual search.

As an example, we do this for object detection; you add & label a few hundred example images and we train a model and stand it up as an API endpoint (and have SDK's for running it locally in a Docker or web browser).

That abstraction & separation of concerns enables hackers to build a computer vision powered app in a couple of hours[1] just like Stripe lets you accept payments in a couple of hours or Auth0 lets you add auth in a couple of hours.

[1] https://www.youtube.com/watch?v=xzh_R8u0hNc

I’ve seen a bunch of drone-detection computer vision projects. Usually they’re detecting drones from other drones though (Eg for autonomous racing[1] or drone-defense).

A challenge with doing it from the ground is that the drones will be quite small relative to the size of the image. But with sufficient compute and several cameras, a tiling-based approach[2] should work.

If you want to do unique-identification you’ll also need object tracking[3].

This is exactly the type of project Roboflow (our startup) is built to empower! Happy to chat/help further (Eg we might be able to help source a good dataset to start from). And if it’s for non-commercial use it should be completely free.

[1] https://blog.roboflow.com/drone-computer-vision-autopilot/

[2] https://blog.roboflow.com/detect-small-objects/

[3] https://blog.roboflow.com/zero-shot-object-tracking/

It doesn't use CLIP but check out NVIDIA's Canvas[1] (powered by GauGAN[2] which has a pretty poor UI but is usable in your browser).

[1] https://www.nvidia.com/en-us/studio/canvas/

[2] http://gaugan.org/gaugan2/

Continually amazed by how versatile CLIP is. I’m keeping a compendium of cool use-cases I run across here[1]. I’ll add this to it today!

[1] https://blog.roboflow.com/openai-clip/

One of the coolest parts about this for me is that Karen built this as a (non-technical) designer.

We took a page out of Twilio's book[1] and made it part of our culture that every new Roboflower builds a computer vision project as part of their onboarding.

As a tool that's meant for non-experts, dogfooding things internally like this surfaces a lot of places where we can make things simpler and more understandable.

[1] https://hbr.org/2016/12/why-we-ask-every-new-employee-to-cod...

Totally. With computer vision especially, you often want to deploy the model on an edge device with limited bandwidth, power, and footprint (think remote camera monitoring, drones, robots, and autonomous vehicles).

Requiring a huge desktop or server-grade graphics card (much less a box full of many of them) to fit the model into memory misses the mark.

We’ve done a lot of work getting models to be performant on the Luxonis OAK (OpenCV AI Kit) and NVIDIA Jetson devices.

Nah this is my my bad.

Revamping the API Key permission logic to be aware that the person who created the project might not be the same person accessing it via API now that we have public sharing was a known “todo” that I deprioritized until users started running into it!

Prioritizing now that there’s actually some traffic discovering it.

I haven’t seen one specifically for parcel delivery vehicles yet. Users have shared package[1] and vehicle[2] projects on Universe.

You could likely use the car detection model as a base to collect photos and bootstrap a collection of delivery truck images. We’re working on enabling users to contribute additional images to these datasets as well.

Creating these custom detectors is what Roboflow is built for (and it’s free if you open source the dataset you create for others to build off of).

[1] https://universe.roboflow.com/joseph-nelson/packages-g0ton [2] https://universe.roboflow.com/roboflow-gw7yv/vehicles-openim...

Uh oh! Is this when trying to get the project info via our REST API (or pip package) for custom training?

I’m working on getting that working with Universe shared projects ASAP but the Jupyter notebook link should work with our training notebooks in the meantime.

Edit: I should mention that inference against the trained api on this project also should work A-Ok already.

Yeah the crop is done after uploading a photo taken towards the general direction of the plane:

> The Function uses a model trained in Azure Custom Vision to detect if there is an airplane in the photo and where it is

Any interest in sharing the plane dataset you used to train Azure Custom Vision with others? Happy to host it on Roboflow Universe along with an API to a trained model others can hit to detect planes in images.

We’ve got one detecting planes on the ground but not yet in the air! https://universe.roboflow.com/gdit/aerial-airport

I’ll keep an eye out

Link to the actual quiz: https://www.buzzfeed.com/daves4/pokemon-ai-quiiz

(I posted Max's tweet because I figured there'd be more relevant info in that thread for the HN audience than Buzzfeed)

Visited the one at Gettysburg a few years ago and it was pretty neat. It's like the 1800s version of virtual reality/3d video.

This makes me nostalgic. The first homebrew app I ever released[1] helped you pick combination locks with the help of your Playstation Portable.

Looks like the original site that described the algorithm is long dead, but the gist was that once you have the first number from a Masterlock (which you can listen for) the search space of the second two can be reduced with a pretty simple algorithm and so there are only a few dozen combinations to try.

[1] http://forums.qj.net/psp-development-forum/7427-release-psp-...

Interestingly, for other systems you sometimes want the exact opposite: for your key space to be distributed across indexes to balance the load (vs wanting them all to hit the same “hot” index for MySQL).

For example, Google’s Cloud Firestore is bottlenecked to 500 writes/second if your key is monotonically increasing (like a timestamp or these timestamp-based UUIDs) causing you to “hotspot” the index: https://cloud.google.com/datastore/docs/best-practices#high_...

If you don’t pronounce the “t” in “often”, how do you pronounce “oft”?

None of this is sacred, and if the team wants to change or eliminate things, we’re open to it and re-evaluate regularly.

This is a tradition that has stayed with the company since the very earliest days when we were in YC. I’ve never heard anyone complain about it, and it’s something I know many on the team (including myself) look forward to each week.

That said, it’s not one size fits all. The culture probably isn’t for everyone; hopefully, being open about how we operate helps people decide whether it’s the ideal environment for them.