Alex Upatov

A bit of product, a lot of communications at TAPe Research Lab. I write about information processing, AI, video search and what breaks outside the demo

[Alex Upatov @hoopatoff]

TAPe+ML v3.5 watching a junction: people, bags, bushes, traffic lights, cars, tables – even people inside the cars.

Being inside another object isn’t much of a disguise.

Model, code and libraries: 4.2 MB altogether.

#ComputerVision #ObjectDetection #TAPeML

[Alex Upatov @hoopatoff]

@ultralytics Great work! Three whole object classes! How much compute, memory and money does all that greatness take?

[Alex Upatov @hoopatoff]

Here’s the photo for anyone who wants to count every last person themselves. Any Rain Men in the house?

[Alex Upatov @hoopatoff]

Here’s v3.5 at work: 1,337 detections of sunlight-deprived humans, plus 296 “not people” objects — including 13 bottles of liquid morale

[Alex Upatov @hoopatoff]

Aye, forgot the bloody hashtags. Used all me brain on the seven Ys.

#ComputerVision #EdgeAI
#MachineLearning #EfficientAI #TinyML #ObjectDetection #InstanceSegmentation #CPUInference #TAPeML

[Alex Upatov @hoopatoff]

A few shitty but pretty self-explanatory pics with tables.

You can read the payper on arXiv right ’ere 👉 https://t.co/jDYpE3jUyy

[Alex Upatov @hoopatoff]

TAPe+ML v3.5. All that, and the whole thing fits in 4.2 MB.

Aye, megabytes. Not a typo.

HNNNNGGGHHHÜÜÜÜÜÜŶŶŶŶŶAAAAARGH

[Alex Upatov @hoopatoff]

and this is the our latest model o my gosh

[Alex Upatov @hoopatoff]

Suzdal. Went out for a pretzel and some lemon-infused vodka. Came back with those, plus 400 photos of churches. I’m a humanities guy. Left unsupervised, this is what happens.

#Suzdal #Russia #Autumn

[Alex Upatov @hoopatoff]

Judging by how this started, the plan is to discuss how machine perception works, check the temperature of the lemon-infused vodka, and somehow end up in Monday.

#TAPe #ComputerVision #MachineLearning #NizhnyNovgorod #Suzdal

[Alex Upatov @hoopatoff]

Celebrating our founder’s birthday at midnight. Came all the way to Nizhny Novgorod to see him, got in at night. Draught beer, dried fish from the market.

[Alex Upatov @hoopatoff]

For breakfast, the math genius (and he is a genius) made excellent shakshuka. We dipped a pretzel I’d brought from Suzdal into it. Suzdal lemon-infused vodka was chilling in the fridge.

[Alex Upatov @hoopatoff]

It reminded me of a client who owns country hotels. He came to us once and asked whether he could train our model to track hares.

Though we also once had a client: a European porn-hosting platform with a library of 5m+ videos.

#ComputerVision #MachineLearning #AI #VideoTech

[Alex Upatov @hoopatoff]

Working from home today – a country house in an empty Russian village.
A warm enough October day to work barefoot on the terrace. Jays are carrying sea buckthorn in their beaks.
Flocks of waxwings, here for the winter, are pecking at apples and pears still hanging on the trees.

[Alex Upatov @hoopatoff]

It wasn’t exactly a piece of cake (probably? I dunno, I’m a humanities guy) – but then again, we had no such product to begin with, and we built one in a couple of weeks.

#3DLiDAR #AutonomousSystems #Robotics #ComputerVision #MachineLearning #MetroTech

[Alex Upatov @hoopatoff]

I’m on a commuter train. Outside the window, it’s pitch-black. Soon I’ll be passing Petushki – a legendary place for anyone who loves Russian letters.

It takes me back to September: we entered a Metro competition and somehow ended up building a 3D LiDAR model while we were at it

[Alex Upatov @hoopatoff]

Tech people don’t dwell on it. They don’t care. Barrels, then.

The prototype saw people in reflective gear at around 120–130 metres, people in black at up to 110 metres, and barrels at up to 90 metres. The smallest objects in the recordings showed up at roughly 45–50 metres.

[Alex Upatov @hoopatoff]

@098uyiubgvgcss The whole thing. Image-to-structure included. Not just a tiny model enjoying a ready-made TAPe representation somewhere offstage. No tricks, just our techno magic.

[Alex Upatov @hoopatoff]

We built a computer vision model with fewer than 100,000 parameters — and on standard benchmarks, it outperforms models that are hundreds or thousands of times larger.

Our current local deployment is just 4.2 MB. And we haven’t even started optimizing it yet.

[Alex Upatov @hoopatoff]

The trial will be time-limited. If it works on your hardware, with your images and your mess, then we can talk about putting it into the product.

So it goes, as Vonnegut used to say.

Elon, knock knock.

[Alex Upatov @hoopatoff]

Yesterday we were discussing what the hell to do next. I was like: we’ve done everything in CV, let bizdev handle it, and point R&D at LLMs and conquering cybersecurity.

But the cybersecurity world takeover has been postponed.

We’re building Edge CV.

#EdgeAI #ComputerVision

[Alex Upatov @hoopatoff]

You install the demo, feed it your own images and video, and see what happens with your light, your shadows, the dirt on your lens and all the places developers usually do not visit.