Send Sequence 2.0

Welcome to the improved Send Sequence action. It now combines all Send action types into one: it can send clicks, type text, move the mouse, and click the mouse, including support for links and anchors.

What’s all the fuss about?

For about 1.5 years, there were 3 separate Send action types: Send Input, Send Text, and Send Sequence.

Each one covered its own specific set of features:

• Send Input was mainly focused on clicking somewhere, such as an image location, anchors, and similar targets.
• Send Text supported several text input methods, such as typing character by character or using the clipboard.
• Send Sequence was always focused on “playing back” longer input sequences, but without the precision of Send Input.

That is now a thing of the past. The redesigned version has the same capabilities as all previous Send action types, and it is much easier to extend with new features and new blocks with any kind of functionality, for example:

• selecting a coordinate
• performing math operations
• using the result to move the mouse

What are the benefits?

The main benefit is that building input sequences is now much easier. For example, if you want to click a specific image, you no longer need to alternate between Send Sequence and Send Input.

I also plan to focus on improving Send Sequence going forward, making it more flexible instead of splitting time across 3 different Send action types.

What should I do with the old Send Input and Send Text actions?

From now on, Send Text and Send Input are effectively in maintenance mode. I will not be adding new options to them or actively working on them anymore.

At some point in the future, they will be fully deprecated and automatically converted into Send Sequence.

You can convert them manually at any time by clicking the Convert to Send Sequence button shown at the top of each Send action. For now, the conversion does not remove the original action, so you can verify that the new one behaves exactly as expected.

Improved image/color/text search: borderless mode

As the title suggests, I added a new option to all of these triggers that allows capturing any window without its border.

In addition, all actions that work by linking/anchoring to the position of a found image will take the missing border into account and adjust coordinates accordingly.

This makes tracking objects in applications that are not running in fullscreen mode easier than ever.

Improved color search

A new option has been added that allows searching for specific pixel blocks of a given color. There is a limited set of supported block sizes, such as 2x2, 4x4, and so on, and the code has been optimized for better overall performance.

Based on my tests, 4x4 blocks give the best results.

Keep in mind that under the hood, this mechanism is very similar to how image search works, so in some scenarios it may still be better to stick with image search. In other cases, though, this new color search mode really shines.

And of course, you can configure anchors/links to the found blocks as well.

Watch the video here

CVATAAT improvements

A fairly detailed documentation page has been added, and it should answer most of your configuration-related questions.

The tool is working and can already be very useful even with its current feature set.

Added support for training with Google Colab

Originally, the tool only allowed training models locally on your own PC using your own hardware. That has obvious limitations: you need to actually have the hardware.

Cloud training makes it possible to use enterprise-grade GPUs such as the A100 for faster and more efficient training in some cases.

It does come with its own limitations, but it can be very useful in many situations. Support is still somewhat limited for now—for example, auto-annotation is not yet available in this mode—but I will continue expanding it.

Bugfixes / Improvements

• [Changelog] Fixed a bug that caused changelogs to be displayed incorrectly, with the text getting “mixed” together
• [Offset, Anchors] Fixed a bug that caused crashes in some cases when adjusting Offset/Anchors in image/text/color search