Post

How to get URL link on X (Twitter) App

On the Twitter thread, click on or icon on the bottom
Click again on or Share Via icon
Click on Copy Link to Tweet
Paste it above and click "Unroll Thread"!
More info at Twitter Help

SkalskiP

@skalskip92

May 20 • 5 tweets • 3 min read • Read on X

I updated my PaliGemma fine-tuning notebook

many of you mentioned that the notebook lacked a benchmark for the fine-tuned model. I have just added mAP and the confusion matrix.

btw, you can expect the PaliGemma fine-tuning tutorial this week

↓ read more

here's the link to the updated notebook: colab.research.google.com/github/roboflo…

I am amazed that in just 3-4 minutes of training, we can achieve a mAP@50:95 of 0.37.

btw, the mAP and confusion matrix calculation is powered by the supervision pip package: github.com/roboflow/super…

here's the conmfusion matrix. not perfect... but still impressive (I think) for just 3-4 minutes of training

btw, here's a blog post explaining what a confusion matrix is and how to calculate it: blog.roboflow.com/what-is-a-conf…

also, if you want to learn computer vision, I maintain a whole repo with notebooks showing how to use and fine-tune different CV models like SAM, YOLOv8, GroundingDINO and of course PaliGemma

link: github.com/roboflow/noteb…

• • •

Missing some Tweet in this thread? You can try to force a refresh

This Thread may be Removed Anytime!

Twitter may remove this content at anytime! Save it as PDF for later use!

More from @skalskip92

SkalskiP

@skalskip92

Mar 25

detecting small objects is hard

I spent some time today writing a short how-to guide on using supervision (in combination with the most popular CV libraries) to detect small objects.

btw is that a good idea for a video tutorial?

link:

↓ read more supervision.roboflow.com/develop/how_to…

this image is 3840x2160; running the model even with increased input resolution (1280) and lowered confidence threshold (0.2) doesn't yield much results

InferenceSlicer processes high-resolution images by dividing them into smaller segments, detecting objects within each, and aggregating the results.

InferenceSlicer docs: supervision.roboflow.com/develop/detect…

Read 4 tweets

SkalskiP

@skalskip92

Mar 21

time analysis with computer vision

- blurring faces
- detection and tracking
- smoothing detections
- filtering detections by zone
- calculating time

let me know if you want me to explain anything else. ;)

code:

↓ read more github.com/roboflow/super…

the full tutorial will be available on Monday on the @roboflow YouTube channel; subscribe so you don't miss it.

link to YouTube: youtube.com/@Roboflow

to ensure the privacy of store employees, I decided to blur their faces; to do this, we will first need a model capable of detecting them.

yesterday, I quickly labeled a few dozen images, which I used to train my model.

this is enough for the demo, but we need many more images to implement such a use case.

dataset link: universe.roboflow.com/roboflow-jvuqo…

Read 9 tweets

SkalskiP

@skalskip92

Mar 19

manual data labeling is (almost) dead

1,500,000 images auto-annotated within 2 weeks of release.

now, we also support automatic segmentation labeling.

↓ read more about open-source models that power this feature

automatic segmentation labeling is powered by GroundedSAM - GroundingDINO + Segment Anything combo;

text → boxes (GroundingDINO)

boxes → masks (Segment Anything)

nearly a year ago, I wrote a blog post about GroundingDINO and its potential use as an auto-annotation tool; now, it's happening.

blog.roboflow.com/grounding-dino…

Read 8 tweets

SkalskiP

@skalskip92

Mar 13

supervision-0.19.0 is just around the corner

among the new features are CSVSink and JSONSink, enabling effortless saving of detections for offline analysis.

I used this feature to create this visualization.

link:

↓ read more github.com/roboflow/super…

https://x.com/skalskip92/status/1743343194727698613?s=20

I used my speed estimation project, which I made a few weeks ago, as a starting point.

link:

https://x.com/skalskip92/status/1743343194727698613?s=20

I decided to visualize the detected vehicles as a 3D matplotlib plot. it's crazy what you can achieve with such simple tools!

here's a comparison of the video frame and its corresponding visualization.

Read 6 tweets

SkalskiP

@skalskip92

Feb 21

The YOLO-World YouTube tutorial is out!

please, let us know what you think!

- model architecture
- processing images and video in Colab
- prompt engineering and detection refinement
- pros and cons of the model

watch here:

↓ more resources

- paper:
- code: arxiv.org/abs/2401.17270
github.com/AILab-CVC/YOLO…

YOLO-World + EfficientSAM @huggingface

space: huggingface.co/spaces/Skalski…

Read 6 tweets

SkalskiP

@skalskip92

Feb 7

defect detection with computer vision

training and deploying manufacturing defect detector step-by-step guide

blog post:

↓ read more blog.roboflow.com/glass-inspecti…