Yoloe -- can we all just go home now?

[Michael Wimble]

Following on from this weeks ROS Sig, let’s take a look at this object recognizer:

[James H Phelan]

I've always thought a robot should be able to ask "Hey boss, what's that?" about something it doesn't recognize.  But, since it knows very little at first, that would be really annoying. So, direct its vision to something you want it to recognize, box it on your screen, type "widget", then say "widget".

How about discriminating very similar objects?  My phone, wife's phone, daughter's phone.  And synonyms. Jim's phone = my phone = black phone.  Linda's phone = wife's phone = pink phone.

Relearning.  It keeps misidentifying something as 'widget'.  Box it.  'Gizmo'.  Maybe just 'not widget'.

With this I can see these things as possible!

James H Phelan
"Nihil est sine ratione cur potius sit quam non sit"
Leibniz

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/035FB3D8-CB9E-4FA8-BE9A-BD4A22279E08%40gmail.com.

[Sergei Grichine]

Selling off my SQUIRRELS stock. HUMANS may be next, but does anyone even care anymore?

Michael, you totally made my day. Interestingly, since I use Ultralytics, swapping the model is a simple one-liner:

https://github.com/slgrobotics/ros2_inference_stereo/blob/main/launch/inference_stereo.launch.py   ('model_path': 'models/yolo11n.pt')

The "prompt training", of course is another hundred lines of code. 

I'll play with it.

Meanwhile....

You are a skilled prompt creator, creating a prompt to visually describe a `squirrel` animal for an AI to visually recognize it. Produce that prompt.

"A small-to-medium-sized arboreal rodent characterized by a hyper-flexible, bushy tail that is often as long as its body and covered in thick, radiating fur. The body is cylindrical and lean, featuring short, muscular limbs with four elongated toes on the front paws and five on the back, all ending in curved, sharp claws for gripping bark. The head is rounded with large, side-mounted black eyes for wide-angle vision and upright, tufted or rounded ears. The fur texture is a mix of dense undercoat and coarse guard hairs, typically appearing in agouti patterns (mottled grey, reddish-brown, or black). Most distinctively, the animal is often seen in a tripod stance, sitting on its haunches while holding a nut or seed with its front paws near its twitching snout."

https://share.google/aimode/fOVJhbR489VL2Mzhf

Best Regards, 
-- Sergei
   

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/d4d9cf58-c014-476b-be4e-35a5e914ff24%40hal-pc.org.

[Sergei Grichine]

Yes, the one-liner change in the launch file works:

  'model_path': 'yoloe-11s-seg-pf.pt', # 'models/yolo11n.pt',  # relative to where you launch, e.g. "~/robot_ws/models/yolo11n.pt"

Without optimization and AI Hat it is noticeably slower though.

I am shopping for AI hat and models for training...

Best Regards,

-- Sergei

[Sergei Grichine]

At least with YOLOE we now know how the squirrels should look like:

https://github.com/slgrobotics/ros2_inference_stereo/blob/main/README.md#fasteroptimized-model-format---onnx

Best Regards,

-- Sergei