Argo Saakyan

[Argo Saakyan @ArgoSaakyan]

@PhysInHistory Hm, mine starts the same way

[Argo Saakyan @ArgoSaakyan]

Everything is reproducible: full 28-model table, per-model configs, protocol, plots.
Benchmark: https://t.co/js2G301sKI
Code (Apache 2.0): https://t.co/U9jLr4Ta8c - pip install dfine-seg
Weights: https://t.co/WNiLY7AkSO - Cityscapes weights are non-commercial

[Argo Saakyan @ArgoSaakyan]

I benchmarked 28 semantic segmentation models on Cityscapes val at native 1024x2048. D-FINE-seg-X came out on top: 84.84 mIoU at 14.6 ms end2end. As far as I know, no real-time model's paper number beats 84.84

[Argo Saakyan @ArgoSaakyan]

Pure TensorRT engine time: YOLO26 wins at every size - 10.4 vs 14.2 ms at X. Deployed pipelines pay end2end (pre + engine + post), gap shrinks: 13.6 vs 14.6 ms. D-FINE-seg S/M/L/X start from an obj2coco pretrain, worth ~+0.7 mIoU at L

[Argo Saakyan @ArgoSaakyan]

@Sentdex That said, api prices are also going down, so we probably will meet somewhere in between

[Argo Saakyan @ArgoSaakyan]

Now you can just pip install dfine-seg!
https://t.co/nILl7Yi7nx

[Argo Saakyan @ArgoSaakyan]

Apple's new M5 ultra in a Mac Studio has 1.2tb/s memory bandwidth and hopefully fast GPU cores. I've been waiting for this machine as you can spec it with a ton of memory and run pretty big models. I think today it is a very decent setup to get 256gb and run Deepseek V4 flash vision with ok prefill (gpu cores) and good decode speeds (memory bandwidth). Not just a decent setup, but something you can do real work with (I've been testing this model alongside claude).

The catch is that it would take like decades to recoup the cost compared to the current API token costs. So yeah, it doesn't make sense from the costs side today, but if you want to play with it and have 10 grand, I wouldn't stop you!

[Argo Saakyan @ArgoSaakyan]

D-FINE-seg now supports 3 tasks: detection, instance segmentation, semantic segmentation. Ahead of YOLO26 and RF-DETR on Cityscapes.
https://t.co/U9jLr4Ta8c

Details - https://t.co/Mdx9wMZiKn

[Computer Vision and Pattern Recognition Papers @CSVisionPapers]

D-FINE-seg: Object Detection and Instance Segmentation Framework with multi-backend deployment

Argo Saakyan, Dmitry Solntsev
https://t.co/bwgaHHIipe [𝚌𝚜.𝙲𝚅]

[Veryfi @Veryfi]

https://t.co/B0pyEKpjdl

[Towards Data Science @TDataScience]

Meet another one of our new contributing authors, @ArgoSaakyan 👋 He is seasoned computer vision engineer, also published on object detection in his recent article, where he examined several well-known pipelines and their customization potential.

https://t.co/tWcMAgMzKD

[Towards Data Science @TDataScience]

.@ArgoSaakyan examines several well-known object detection pipelines and designs one that best suits his needs and tasks. Learn how to master object detection, read the full article free now.

https://t.co/tWcMAgM1V5

[Towards Data Science @TDataScience]

Build your own object detection training pipeline. @ArgoSaakyan's new article provides a step-by-step guide to creating a custom pipeline, with insights into data augmentation, training optimization, and model evaluation.

[Argo Saakyan @ArgoSaakyan]

I found out that a pretty nice technique in training DL models is rarely used or mentioned. It’s Exponential Moving Average in model’s wights updating.

I implemented it in my detection pipeline and model converged significantly faster with better scores

[Argo Saakyan @ArgoSaakyan]

If you don't log/monitor your DL model experiments or you don't have your CUDA variables fixed, I recommend reading my recent article about running experiments with PyTorch, WandB, Hydra and other tools. It makes life easier.

https://t.co/ZyTvUcmLiQ

[Argo Saakyan @ArgoSaakyan]

@elonmusk Will DOGE deal with year long administrative processings for O1 (and similar) visas? Or is it out of the scope of the new department?

[Argo Saakyan @ArgoSaakyan]

Hey, I am a Computer Vision Engineer/Researcher, I work at Californian startup and I write articles, do some mentoring, judge hackathons/awards. Will be posting some insights (Deep Learning related) and links here.
https://t.co/T7XYMSOSVi