Pixel perfect masks are the most accurate type of label that you can have in 2D computer vision but also the most expensive to create. The average time it takes to complete a polygon mask annotation is about 34 seconds across the industry. Auto annotate helps you reduce that to about 2. 5 seconds.
What this ultimately means is that in five hours, you're able to create almost 10, 000 labels. This makes it as fast as creating bounding boxes without the mental fatigue of having to perfectly align crosshairs and edges. It's able to learn anything from these tools on this image. to clustered objects like the ones that you're looking at now.
It doesn't matter what angle they have, what lighting conditions you're looking at, and it's not just simply a statistical type of hackery. It's actually using a deep learning model that is trained on such a large variety of items. That it's now able to segment almost everything. Why do we recommend using this tool?
Firstly, it's faster. This is easy to measure, and it will save you between 65% and 90% of your time spent labeling. Secondly, it reduces your mental fatigue a lot. One of the most difficult things about labeling images is not necessarily the act of clicking, it's the mental fatigue that comes from accurately clicking around pixels for hours every day.
Finally, it's pixel perfect. Humans aren't that good at following pixel gradients exactly. It's a tedious job, and it's one best left to machines. What Auto Annotate helps you do is simply define a rough bounding box around each object. You don't really have to align the crosshairs exactly to the corners, you can simply create a very broad region of interest, and then it will understand the most salient object within that boundary.
If you zoom into one of those output masks, you notice that the border is exactly mid gradient. This gives the best result when training an object, and Auto Annotate helps you do this effortlessly rather than having to manually move the borders of each mask. Even if objects overlap such as these x rays, it's still able to understand what object you intended to focus on.
Another difficult case, for example, is people in difficult lighting or behind smoke. Auto Annotate is able to cover this case perfectly, which is something that SuperPixels or traditional computer version methods would never be able to achieve. It might make a mistake, and you can simply click to correct it immediately.
If you want an object to be expanded beyond what it seems to be the most reasonable item within that region, Keep clicking outside of the object and Auto Annotate will rerun the same model to include those new parts. It's meant to work hand in hand with you as you label these images or have them labeled by your workforce of choice.
If you're labeling images yourself, this is both a great time saver and a sanity saver. And if you have a workforce currently labeling images, it's hopefully a huge money saver as well. Auto Annotate is available within any image domain including medical and pathology images. any image resolution, file type, or size, including video frames.
We are also able to fine tune it to specific data sets once you have some training data should your objects be particularly strange to the point where Auto Annotate doesn't fit to them perfectly on the first go. Auto Annotate currently covers the vast majority of pixel masks done on V7 Darwin and we hope you'll get a chance to try it out as well and complete your data sets 90% faster and more accurately.