Generate thumbnails using smart crop sometimes fail miserably when using eg animals

[marcuslindblom]

We have this media library and I want to create good looking thumbnails for most of the images that are uploaded. I have activated advanced smart crop in my container and this is the parameters I'm using /w:240/h:135/dpr:2/rt:fill/q:35/gravity:sm/

The result with this source looks like this.

This view is from our library and it does not do a good job. The first image has a car in the picture but that is not visible at all. Is it possible to fix this using some other parameters or do you have any other suggestions?

[DarthSim]

Both smart crop and advanced smart crop are not content-aware. They both work on the pixel level using different techniques. This means that they don't see any difference between a car, a tree, a fox, or a wall. The only thing that matters is if the pixels meet the requirements. For example, here's the features map of the fox photo:

As you can see, the asphalt under the fox's paws is noisy, so it catches the detector's attention. Such algorithms are doing a good job when you're dealing with images containing objects on a quite solid background or photos with good bokeh. But when the entropy of a whole image is uniform, such algorithms will most probably fail.

If you want a content-aware crop, you should probably take a look at the object detection feature. Though the default model that is delivered with imgproxy Pro is trained only on faces, you can use any pre-trained DarkNet YOLOv4 model with imgproxy or you can train your own. Drop us an email if you need any help with object detection.