Opinion on my annotations

Show & Tell
1

Hello there, I’m developing a model that is able to detect boxes of two types, mainly for accademic purposes.
I was wondering my annotation skills and i’d like an opinion on how’s my work and eventual tips and tricks for the best annotations in order to better train my models.

Here’s my dataset:

Thank you!

2

Hi @Damiano_Purin!
Welcome to the Roboflow community!

Do you mind providing screenshots of your annotations? I was unable to access your dataset via that link. Thank you!

3

hey, thanks for the reply, that’s wierd my project should be public, anyways here are some images:

image
image778×783 48.1 KB

This one is wierd because i was trying my best not to include the white box with the other brown one and also avoid including parts of boxes that wasn’t the “target” one

4

image
image943×933 58.3 KB

also this one is a bit wierd, i didn’t know what to include

5

image
image941×931 75.8 KB

Am i includin too mutch background here? That is also a thing i’m confused about, cuz it happends sometimes that i don’t know how to annotate with sqares things that are rotated that mutch

6

image
image938×918 49 KB

So i tried using the smart poligon, but i couldn’t understand if this was good enough

Thanks for your time

7

I was also wondering if I should use the polygon tool even if i’m looking for making an object detection model, i know it will then simplify the polygons into boxes, but i was thinking that maybe this way i would solve the “background problem”

8

Hi @Damiano_Purin!
Great question! I will comment on each individually.

Here is a great blog post from our Founder explaining labelling data further: How to Label Image Data for Computer Vision Models

9

It is ok to have annotations overlap, its most important to get the entire object of interest (the box) inside of the bounding box.

10

Overlap is ok, its most important to have the boxes on the right within the bounding boxes.

11

This is perfect!! Nice job

12

Also looks great here as well!!

13

Thank you @Ford, I have 2 last questions:
The first is, should i convert all the polygon annotations back to boxes, since i’m training an object detection and i’m not interested for a segmentation model, i feel like it’s better to turn them back to boxes.

The second is, is it ok to include background as long as the box is as tight as possible around the object?
For example in this image, since the objects are rotated, the bounding boxes need to include a bunch of background, is it ok?

image
image834×830 73.5 KB

Thank you a lot!

14

Hi @Damiano_Purin!
No worries! You can leave them as polygon annotations.

This is ok! It is a tradeoff that must be accepted because we need the entire object of interest within the bounding box. You can always implement polygons to mitigate this.

Happy building!