Roboflow
2.1K posts
Roboflow
@roboflow
Build & use computer vision models fast ✨ Get started: https://t.co/QvAsWtf7c8 Open source datasets & models https://t.co/IzwgoIkBEe
Katılım Haziran 2019
1 Takip Edilen13.8K Takipçiler
@roboflow At the moment, LLMs are extremely poor when it comes to computer vision and object detection.
They can do basic things right sometimes, but for complex detections, its still very basic and amateur.
But it will improve rapidly.
English
looks like grok is entering the vision arena soon
cool to see elon in the weeds on that
SkalskiP@skalskip92
GPT 5.6 Sol is the best "vision" model OpenAI ever released massive gains in object detection and counting. still very strong in OCR ↓ deep GPT 5.6 dive
English
related?
x.com/Techmeme/statu…
Techmeme@Techmeme
Sources: Cursor is building a general-purpose AI agent codenamed Sand, aimed at non-developers, that handles emails, texts, and documents to rival Claude Cowork (@graceihle / The Information) (Visit Techmeme dot com for the link and full context!)
English
Perfect 👌 The bounding box text needs different colours, but detection is accurate despite this wild-angle screen capture. Drone is sitting on the chair 😬
On Saturday we’ll test the final app with an elite soccer school. We’ll stream live on YouTube with AI commentary 🤞
Sunday: go to market with the new page.
Monday: 95k emails to my sports network in Germany.
4 weeks from first demo to final product. Thanks to @claudeai & @GeminiApp / @GoogleAIStudio / @Lovable / @roboflow / @tripoai 🙏
✅ Broadcasting with AI commentary
✅ Multicam
✅ Multi-stream (3x)
✅ Infinite ♾️ CV model loading
✅ Tactical AI
✅ All exercises
✅ Anatomy tricorder
✅ sportscience.md
✅ AR / VR
✅ Sensor (8x) fusion
✅ iOS / web
✅ Stats shared to Booca Scouting Hub @BoocaMart
✅ So much more
The best part: this app is the foundation for my construction use cases. I just have to switch the CV models and .md
Btw: we’re searching for ambassadors (pro athletes, sport influencers).
DM if you’re a sport agent to negotiate the fee.
M. 🫡
Miles Dyson@menrva33
Just added a real FLIR thermal cam to the app. 2 streams, same server. CV running on both screens. Thinking about sensor fusion as a little helper for the CV models. Waiting on the 3D-printed case now. Can’t wait to test it in flight. Only cam missing now: 360°. Anyone got a 360° cam recommendation with WLAN/streaming? Anyway - roofing guys and everyone else from my construction bubble will save serious money if the accuracy holds up. 🤞 Why spend 7k–9k when you can have it for under 1k and with CV… Logic: same as always…you can achieve 80% of results with 20% of the costs. #allinone #drone #construction #computervision #macgyver
English
@roboflow At least a “thank you” would be nice guys
English
Roboflow retweetledi
object counting is one of the most popular use cases for VLMs
what previously required training a model on a custom dataset now fits into a prompt
quality control, inventory checks, shelf monitoring
check how different VLMs handle object counting: playground.roboflow.com/share/80153ee6…
English
Roboflow retweetledi
I’m evaluating fable 5 on object detection, but so far it performs much worse than gemini 3.5 flash
any tips?
it struggles with cluttered scenes and overlapping objects
try gemini 3.5 flash vs fable 5 for free: playground.roboflow.com/share/807336ce…
↓ some positive fable 5 examples
English
very good thread with sample data showing Fable 5 performance against other models for vision tasks
important to see the types of tasks frontier models can/cant do
will give you ideas for how to use them successfully in production
SkalskiP@skalskip92
need a model for OCR and VQA? Fable 5 is NOT your best option Gemini 3.5 Flash is 1.3x faster, 2.9x cheaper, and more accurate Fable 5 struggles especially on tasks that require counting objects. ↓ cost comparison and interesting failure cases
English
Hey @roboflow
I don’t usually do this, but this role stuck with me.
You’re hiring a Technical Recruiter and I’ve spent the last few years working closely with teams like 0x, Matter Labs and Protocol labs building hiring systems for technical roles in AI/infra environments.
I’ve also actually used @roboflow in practice for over a year, which I suspect isn’t that common from the recruiting side.
Most of my work has been around figuring out how to find strong candidates that don’t show up in normal pipelines, and helping teams understand what “good” actually looks like beyond interviews.
Feels very aligned with how you think about hiring.
If anyone on the team is open to chatting, I’d really love to connect.
English
@leon_cvetkovski 3.5 35B is the #2 model, we tested the 3.6 version and it didn't improve results
English
@roboflow When are you going to test the 3.6 35b model, super interested for the results
English
Fable 5 is back and tied as a top vision model, Sonnet 5 is outside the top 10
Fable 5 is ~3x slower and ~10x more expensive than Gemini 3.5 though
If you're set on Anthropic models, it's the best one for vision
The delta between Sonnet 5 and Fable 5 is really large, Sonnet 5 is slightly worse than Qwen 3.5 9B
but very fast if you're looking for speed!
Roboflow@roboflow
You can now build vision applications with Fable 5 in Roboflow Fable 5 is the best Anthropic model for visual understanding, beating Opus 4.8 on our evals very high object understanding capability means it's a great option when using a VLM as a judge in your pipeline ideal for a final step in a process that requires judgement or interpretation of a situation of course, nails most document, label extraction, ocr, scene reports, etc that you expect of a large frontier model
English
link to run your own images against all threes models to see how they compare
playground.roboflow.com/models/compare…
English
every legacy vision product/system/company can probably be built with open source tools and models at this point
and then putting them in production is 100x cheaper than it's ever been
vision is no longer cost prohibitive to build and deploy
mirrash.eth 🦇🔊@mirrash7
I rebuilt FIFA's VAR offside tech run on a single camera for about $2 a match 📈 -Find pitch lines + vanishing point- HSV color matching -Find players - RF-DETR -Recreate the moment in 3D - SAM3D Body + 3js
English
We added video understanding and Gemini 3.5 Flash wins there as well
Seems like the all around best frontier vision model
If you're starting to build vision apps, use Gemini 3.5 Flash to prototype
You can quickly build a work pipeline across different use cases and types of data, then move to a fine-tuned model as you scale up
Expanding the video benchmark soon, just getting started
Roboflow@roboflow
we saw the same thing with our vision leaderboard still testing and wrapping our heads around Gemini 3.5 Flash but it feels like a pivotal shift in frontier models way cheaper and way faster than anything else, and performance is super high first time it feels like you can really use a VLM for high scale tasks. lots of previous tasks that required making workflows of smaller, cheaper models now looks possible (speed and cost wise) with Gemini 3.5 Flash a few examples below
English
Roboflow retweetledi
just published a step by step guide for fine tuning RF-DETR keypoints
pretrained checkpoint does coco 17 point human pose
I fine-tuned it on basketball court keypoint dataset for 33 court landmarks
repo: github.com/roboflow/rf-de…
blog: blog.roboflow.com/train-rf-detr-…
English
@cmaye35 yeah here's an example of that
blog.roboflow.com/american-footb…
English
