Roboflow Blog

Roboflow | Solutions Architects, Implementation Engineer, Developer Advocate, Full Stack Engineers | Full-time (Remote, SF, NYC) | https://roboflow.com/careers?ref=whoishiring0725 Roboflow is the fastest way to use computer vision in production. We help developers give their software the sense of sight. Our end-to-end platform [1] provides tooling for image collection, annotation, dataset exploration and curation, training, and deployment. Over 1 million engineers (including engineers from 2/3 Fortune 100 companies) build with Roboflow. We now host the largest collection of open source computer vision datasets and pre-trained models [2]. We are pushing forward the CV ecosystem with open source projects like RF-DETR [3] and Supervision [4]. And we've built one of the most comprehensive resources for software engineers to learn to use computer vision with our popular blog [5] and YouTube channel [6]. We have several openings available but are primarily looking for strong technical generalists who want to help us democratize computer vision and like to wear many hats and have an outsized impact. Our engineering culture is built on a foundation of autonomy & we don't consider an engineer fully ramped until they can "choose their own loss function". At Roboflow, engineers aren't just responsible for building things but also for helping us figure out what we should build next. We're builders & problem solvers; not just coders. (For this reason we also especially love hiring…

Roboflow | Multiple Roles | Full-time (SF, NYC, Distributed) | https://roboflow.com/careers?ref=whoishiring1023 Roboflow is the fastest way to use computer vision in production. We help developers give their software the sense of sight. Our hosted tools are used for image/video collection and annotation; dataset exploration with foundation models; model training; deployment to anywhere. Here's a simple webapp I made recently: https://josephofiowa.com/rps-embedded-vision23/ Over 250k engineers (including engineers from 1/2 Fortune 100 companies) build with Roboflow. We now host the largest collection of open source computer vision datasets and pre-trained models[2]. We are pushing forward the CV ecosystem with open source projects like Autodistill[3] and Supervision[4]. And we've built one of the most comprehensive resources for software engineers to learn to use computer vision with our popular blog[5] and YouTube channel[6]. We have several openings available, but are primarily looking for strong technical generalists. Our engineering culture is built on a foundation of autonomy & we don't consider an engineer fully ramped until they can "choose their own loss function." At Roboflow, engineers aren't just responsible for building things but also for helping figure out what we should build next. We're builders & problem solvers; not just coders. (For this reason we also especially love hiring founders.) We're currently hiring full-stack engineers for our ML and web…

Roboflow | Multiple Roles | Full-time (Remote, SF, NYC) | https://roboflow.com/careers?ref=whoishiring0923 Roboflow is the fastest way to use computer vision in production. We help developers give their software the sense of sight. Our end-to-end platform[1] provides tooling for image collection, annotation, dataset exploration and curation, training, and deployment. Over 250k engineers (including engineers from 2/3 Fortune 100 companies) build with Roboflow. We now host the largest collection of open source computer vision datasets and pre-trained models[2]. We are pushing forward the CV ecosystem with open source projects like Autodistill[3] and Supervision[4]. And we've built one of the most comprehensive resources for software engineers to learn to use computer vision with our popular blog[5] and YouTube channel[6]. We have several openings available, but are primarily looking for strong technical generalists who want to help us democratize computer vision and like to wear many hats and have an outsized impact. Our engineering culture is built on a foundation of autonomy & we don't consider an engineer fully ramped until they can "choose their own loss function". At Roboflow, engineers aren't just responsible for building things but also for helping figure out what we should build next. We're builders & problem solvers; not just coders. (For this reason we also especially love hiring past and future founders.) We're currently hiring full-stack engineers for our ML…

Hi, I’m the author of this blog. It depends on how strong of a model you need, but in general, Qwen is easily the best among the Chinese models right now. Over the last two weeks, Qwen released two new models. Qwen3.8-Max is totally insane, but it’s only available through the Alibaba Cloud API. I wrote a similar blog covering Qwen3.8-Max: [https://blog.roboflow.com/qwen3-8-max/](https://blog.roboflow.com/qwen… If you’re looking for something you can run locally, Qwen3.8-27B might be a great option. On Friday, I did a quick comparison between Qwen3.8-Max and Qwen3.8-27B: [https://x.com/skalskip92/status/2088411215441621469?s=20](https://x.co…

I'd assume they'll use something like https://blog.roboflow.com/rf-detr-segmentation-preview/ and then use the "classes" as embeddings for your profile to do a similiarity match. Or the standard given a bunch of tags (from the instance segmentation) and find your ranked list of preferences. https://librecommender.readthedocs.io/en/latest/ has a list of 5-10 standard recommender algorithms. Or use autogluon https://auto.gluon.ai/stable/tutorials/tabular/tabular-multimodal.html where they take Meta's large repository of annotated-images (DINOv2) and use them to classify then do the recommendation system on tabular (database tables). This stuff has been standard for years now.

Keep in mind that Apple's claimed false positive rate (one in a trillion chance of an account being flagged innocently), and the collision rate determined by Dwyer in the blog post linked from the repo [2], are both derived without making any adversarial assumptions. Given that NeuralHash collider and similar tools already exist, the practical false positive rate is now expected to be much much higher. Imagine that you play a game of craps against an online casino. The casino throws a virtual six-sided die, secretly generated using Microsoft Excel's random number generator. Your job is to predict the result. If you manage to predict the result 100 times in a row, you win and the casino will pay you $1000000000000 (one trillion dollars). If you ever fail to predict the result of a throw, the game is over, you lose and you pay the casino $1 (one dollar). A casino that makes no adversarial assumptions about the clientele could argue as follows: the probability that you accidentally win the game is much less than one in one trillion, so this game is very safe, and the House Edge is excellent [3]. But this number is very misleading: it's based on naive assumptions that are completely meaningless in an adversarial context. Some of the clientele will cheat. If your adversary has a decent knowledge of mathematics at the high school level, the serial correlation in Excel's generator comes into play [4], and the relevant probability is no longer less than 1/1000000000000. In fact,…

https://blog.roboflow.com/neuralhash-collision/ Seems like about 3 collisions per 100 million pictures. If everyone have 1000 pictures that is 3 collisions per 100 000 users. Raiding 2 people in my city for made up CSAM pictures would be way to high false positive rate. But you can also make any picture match a CSAM hash by adding picked noise.

From the website: >NYCerebro is a CLIP-powered search engine for NYC traffic cameras. It uses AI to find camera views matching your text descriptions. Blog post: https://blog.roboflow.com/nycerebro/ Github: https://github.com/yeldarby/nycerebro

We got bboxes from yolo indeed to identify "here is a traffic sign", "here is a traffic light" etc. Then we cropped out these objects of interest and took the DINOv2 embeddings of them. Not using it to create segmentations (there are YOLO models that do that, so if you need a segmentation you can get it in one pass), no, just to get a single vector representing each crop. Our goal was not only to know "this is a traffic sign", but also do multilabel classification like "has graffiti", "has deformations", "shows decoloration" etc. If you store those it becomes pretty trivial (and hella fast) to pass these off to a bunch of data scientists so they can let loose all the classifiers in sklearn on that. See [1] for a substantially similar example. [1] https://blog.roboflow.com/how-to-classify-images-with-dinov2

This is not the official announcement. It's from some blog with horrible picture quality: https://blog.roboflow.com/content/images/2024/05/image-11.png

Internally that is probably something similar that is happening, but as far as I can tell it was never designed to specifically do that. Here is an example of the things GPT-4V is capable of: https://blog.roboflow.com/gpt-4-vision/amp/

FWIW, GPT4V (which is what I assume Bing uses behind the scenes) performs considerably worse on Recaptcha[1]. [1] https://blog.roboflow.com/gpt-4-vision/

There is a similar project for rabbit deterrence that might give some helpful ideas on the approach: https://blog.roboflow.com/rabbit-deterrence-system/

Elsewhere in this subthread was mentions CLIP[0] About 2/3rds down the page features 'Object tracking' and 'Robotic control' [0] https://blog.roboflow.com/openai-clip/

I think the Roboflow tool does address a lot of that too, though. You can sample classes with lower confidence detections (and at specified IOU thresholds) and bring them back in as new training data. https://blog.roboflow.com/upload-api/ If you were to combine the "smart-sampling" with things like the dataset health check for more insight into the class balance, object count histograms for the labeled data, etc. would that be a good first step in curating/developing the better training sets?

This is a computer vision model I made to see if I could use computer vision to help me identify the chocolates in a box of See's Candies. You can try it out at https://universe.roboflow.com/chocolates/valentines-chocolates If you're interested in reading more about the project and how I did it, I wrote a post for the Roboflow blog: https://blog.roboflow.com/identifying-chocolates-with-computer-vision/

I'm somewhat surprised neither this article nor the previous one mention anything about the Florence-2 model series. I had thought that Florence-2 was not just surprisingly capable for this kind of work, but also easily fine-tunable for a particular kind of document, when you expect to process a lot of instances of that document and want to further optimize accuracy. It's extremely small (0.23B and 0.77B parameters), so it's easy to run, easy to fine-tune, and probably unlikely to overthink things. https://arxiv.org/abs/2311.06242 https://huggingface.co/blog/finetune-florence2 https://blog.roboflow.com/florence-2-ocr/ https://www.assemblyai.com/blog/florence-2-how-it-works-how-to-use/ I don't personally deal with any OCR tasks, so maybe I misread the room, but it sounded promising, and I have seen some continuing interest in it online elsewhere. In addition to the architectural issues mentioned in OP's article that are faced by most SOTA LLMs, I also expect that current SOTA LLMs like Gemini 2.0 Flash aren't being trained with very many document OCR examples... for now, it seems like the kind of thing that could benefit from fine-tuning on that objective, which would help emphasize to the model that it doesn't need to try to solve any equations or be helpful in any smart way.

"Apple said that there is a one in one trillion chance of a false positive" https://techcrunch.com/2021/08/05/apple-icloud-photos-scanning/ Versus: https://blog.roboflow.com/nerualhash-collision/ Along with: https://github.com/anishathalye/neural-hash-collider Demonstrates blatant incompetence on the part of Apple's design. At this point, trust is broken. That's the end of credibility regarding any other claim made about the security or base function of the system. As for a simple example of one of many potential attack vectors (search HN for "neuralhash" and read the comments for countless examples), here is a basic attack: As Apple's hash algorithm is entirely broken and useless, a malicious actor can easily craft a NSFW, but otherwise legal image, that has a hash matching one in the CSAM database, that is visually ambiguous in that context (for example, a close-up). Targeted attacks could be launched on unsuspecting victims as simply as sending them an image. If their app is configured to auto-save to iCloud - as some are by default, and as many are configurable to be - the image will then get automatically flagged as a match, and may well potentially pass a human review - because the image may indeed look just like CP - and an innocent person, who may not even be aware of the presence of the image in their iCloud library at this point, may get visited by the police. That in itself is bad enough. However it gets worse - do bear in mind that "the police" is not…