Tracking Custom Objects - TensorFlow Object Detection API Tutorial p.3

Welcome to part 3 of the TensorFlow Object Detection API tutorial series. In this part and the subsequent few, we're going to cover how we can track and detect our own custom objects with this API.

Going from using the pre-built models to adding custom objects is a decent jump from my findings, and I could not locate any full step-by-step guides, so hopefully I can save you all from the struggle. Once solved, the ability to train for any custom object you can think of (and create data for) is an awesome skill to have.

[planktonfun]

shortcut keys
w -Create a rect box
d -Next image
a -Previous image

[vivekkalyanarangan]

There is a dope chrome plugin called Fatkun that batch downloads image (you have the option of selecting which ones as well). Hope that helps!

[justchill]

As @Paulo Castro rightly said,
Shortcut keys
w -Create a rect box
d -Next image
a -Previous image

and from me :
In the LabelImg window, Go to View and then check Autosaving. BOOM! Now you can draw a box and then just press 'd' for the next image, it will automatically save the .xml files(first time it will ask the destination folder though).

PS: Another awesome speed up tip - If you are just using one class, look on the top right side. Check the option 'Use default label' and enter the one class name you want to assign. For example, 'car' if you want to detect car in every image.
This way it will not ask you the label name everytime you draw a RectBox. So just draw the RectBox around the car and press 'd' for next image. This automatically saves the generated .xml file and with the one and only label you want to assign! So useful !!!

Cheers!

[subramanyam]

There is a very very good reason to train our own custom model. I wish there was a tutorial on that..

[hamedatohamedato]

Folks! Installing labelImg is not difficult at all and you don't have to switch to Ubuntu if don't want to. Maybe it was difficult in past or maybe it's still difficult if you follow Github's installation instruction but there's an easy way to install it. Unlike Github's installation instruction that says to go to the link for PyQT5 and install .exe and before that you would also need to install Win10 SDK for Microsoft VS, Sip and QT, which are really pain in the butt and at the end QTCore won't be recognized, just install it through pip: pip install pyqt5. That's it guys! You saved hours of installation of different software and packages! Then, again there's a mistake in Github's installation instruction, so type this in a cmd opened in your labelImg directory:
pyrcc5 -o libs/resources.py resources.qrc
And finally:
python labelImg.py
And life is good afterwards!!

[cryptomustache]

Mining those crypto currencies like mad. Nice job bro. Knowledge == Money. This era is the best one, for really smart people. Thanks to you I become a millionaire. I wrote you 2 years ago on e-mail. Thanks for everything. Angels like you make this world a better place.

Keep rocking.

[yahavelt]

Really enjoying the series. Plus I was really interested in the TensorFlow Object Detection API!!!

[kaushikprakashrobotics]

Thank you so much for these tutorials! It's amazing how much I can understand from your videos.

[tonifaunuscloet]

Nice to have tutorials coming frequently!

[sairam]

Thanks. These videos saved a ton of my time

[shreeyaksajjan]

Hey, I'm loving this tutorial series, thanks so much! I was trying myself to train a system to detect my custom images, I think I used the system trained on ImageNet that does not draw any bounding boxes, so this is AWESOME!

It would be nice if you could put the links of all the sites used in the video in your description. makes it much nicer. Like the image labelling tool you used, for example.

Great work, keep it coming!

[mehmetsoydas]

@sentdex hello thank you for the tutorials, i wonder do you know how data augmentation is done with ssd method? I mean it has to have xml files of all images. If i want to augment my data with flipping horizontally and mirroring, will i have to label all augmented data as before we did? Think about i will do augmentation with flipping all images 15 degree horizontally and mirror them both horizontal and vertical. This means my data will be 96 times augmented and labeling all of them will be so tiring process :D what will be your suggestion for augmentation, are there any easy way to do this? Thanks again.

[OmegAtlAnt]

How are you always in such a good mood?

[sohommukherjee]

Thanks a ton sentdex! Your turorials really help me so much...

[kerwinsibunga]

Is it possible to remove other label in the pre trained model without retraining the model?

[suee]

Sentdex great tutorial!, just curious on the training part, what if I wish to add another class let say barbecue, do I need to re-train again including the macaroni and cheese or is there any way to append the already trained data?

[riyabanerjee]

Hello,

Can I use the above program for Tensorflow CPU support?

When I run the code (video 6) I get the following error:

File
thon/framework/errors_impl.py", line 519, in __exist__

=; Is a directory

I do not know if this error is because I use Tensorflow CPU or some other reason

[rjk]

Can you use fewer images, but with multiple copies of the object? 50 images containing 2 objects each is the same as 100 images containing only 1 object each, right?

[simonw.]

At 3:52, I knew there was no way you were going to say macaroni and cheese, given the thumbnails of the other videos in this playlist. I was wrong. You said macaroni and cheese. 😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂

[JoshStepp]

To save cost... couldn't you do all the dataset and data cleaning offline and SCP or pull the data to paperspace? Like pull all images, label them as macnchesse and the transfer them to you on machine? Or am I missing something?