open source @roboflow | AI engineering at @UdeSA🇦🇷 prev founder @SatsOnFire
Joined May 2016
- Tweets3.9K
- Following1.1K
- Followers2K
- Likes19.5K
Pinned Tweet
I’m thrilled to present the KAN Convolutional Layer, a promising neural architecture for image processing. Last week KANs came out, as an alternative to the MLP. We extended this idea to Convolutional Layers, creating the KAN Convolution. Join me in this thread to more about it🧵
Terminando la semana de puentes con la personal Agents Hackathon by @AntigravitySF en las tremendas oficinas de @firstround
A ver cómo le va al team @chiaraciriani_ @anaundu @MartinPulitano
Incredible day at @roboflow VIS.
Met @skalskip92 after working together for over a year, discussed OpenCV and open-source’s present and future with @LearnOpenCV , heard SAM authors (one of the most important works on CV of last years) and definitely felt that the world will be shaped by people in that event.
Alex Bodner retweeted
Puentes starts in two days!
If you want to meet any fellows, request intros here:
puentes.antigravity.capital/…
Puentes starts this Sunday.
We're flying 15 of the best young engineers from Latin America to SF.
They’re among the most talented in their countries, and they want to be here to compete with the best.
Our goal at @antigravitysf is to help them stay.
Meet the engineers 👇
How do we do to track an object accurately with a camera moving? Meet Camera Motion Compensation!
In practice, you can use BoT-SORT in trackers with a 1 line change:
github.com/roboflow/trackers
Object tracking explained in 3Blue1Brown style.
Learn about the tracking loop, kalman filters, why occlusion breaks and one way to fix it.
Try it out github.com/roboflow/trackers
Alex Bodner retweeted
Gemini 4 Argon stream in 30 minutes. see you!
Gemini 4 Argon vs GPT-6 Astra YouTube stream
tomorrow, Oct 15th, 8 AM PT / 11 AM ET / 5 PM CET
save the time and see you tomorrow
- detection
- segmentation
- transcription
- data extraction
- efficiency and cost
what else?
link: youtube.com/live/bBX_c5AFuMQ
Alex Bodner retweeted
The fellows want to meet engineers at:
@SierraPlatform
@posthog
@chaidiscovery
@linear
@WisprFlow
@AnthropicAI
@OpenAI
Instinct
@SpaceXSI
@modal
@PrimeIntellect
@ssi
@theworldlabs
@beltichq
If you work or know anyone there, dm for dinner invite :)
Puentes starts this Sunday.
We're flying 15 of the best young engineers from Latin America to SF.
They’re among the most talented in their countries, and they want to be here to compete with the best.
Our goal at @antigravitysf is to help them stay.
Meet the engineers 👇
Alex Bodner retweeted
Puentes starts this Sunday.
We're flying 15 of the best young engineers from Latin America to SF.
They’re among the most talented in their countries, and they want to be here to compete with the best.
Our goal at @antigravitysf is to help them stay.
Meet the engineers 👇
Muy contento de participar en esta edición de Puentes! Voy a estar viajando el viernes para SF, y buscando conocer gente del ecosistema. Si trabajas en AI, vision, robotica, NLP escribime!
@AlexBodner_ (22, Argentina). At 16, he co-founded Satellites On Fire and built its wildfire models and alerts (900 monthly users). He now builds Roboflow’s object tracking library (380K+ downloads, 3.9K+ stars) and is first author of Convolutional KANs (170+ citations).
My favourite thing about @roboflow Trackers is how easy we made it to swap things and create the best tracker for your case.
Swap the tracker, the IoU variant, camera motion compensation, dynamic frame rate or masks, one line change, and see what works best!
Any detector works.
Find more in:
github.com/roboflow/trackers
Leave a star if you find it useful, it motivates us🌟
I'm still working on the AutoTracking dataset creation. Now I've tried GPT-6.1 Sol on decisions a human would need to make.
Astra was good but expensive. At 1/4 of the cost, Sol performs about the same.
Stay tuned
We are trying to give a model three words and get back a labelled tracking dataset. No human labels anywhere.
After a week, its still not perfect, but the self improvement is working. Out of Astra + SAM3 we can generate some tracks that fragment, but we can use that for training cheaper models that allow us to detect players better and join tracks.
Numbers so far:
On SoccerNet's official test split, 46 clips it never saw: 62.5 HOTA.
For reference on the same clips, published baselines that use SoccerNet's labelled train split or pedestrian datasets:
- DeepSORT 36.7
- ByteTrack 47.2
- FairMOT fine-tuned on SoccerNet 57.9
We use none of those labels. Still far behind the 2023 challenge winner at 75.6.
Excited for next week progress👀
For which use case would you use this?