OPHoperHPO
R-4.5.0
Changelog:
- Removed obsolete models
- Fixed typos in tracer b7
- Added new AutoScene preprocessing method
- Added warning about fp16 to u2net class
- Updated README.
- Changelog
- Added ISNet segmentation network #119
- Added object classifier YOLOv4.
- Added AutoInterface to automatically select the best methods for images with different scene types. AutoInterface uses a scene classifier and an object classifier to perform full analysis on all images. At the moment, since there are not enough models for so many types of scenes, universal models are selected for some domains. In the future, when some variety of models is added, auto-selection will be rewritten for the better.
- Updated metrics
- Added CascadePSP segmentation refinement network. It will refine the segmentation mask before passing it to the matting network. This should provide more accurate background removal for objects and people. (Check repo from CREDITS.md for more details)
- Fixed trimap generator
unknownarea value as FBA Matting expect. - Added noise filter switch (needed for smooth mask prediction on the final stage for
soft scenes) - Updated alpha composition algorithm
AutoScene RC-4.2.0
Description
It performs a primary analysis of the image in order to automatically select the necessary method for removing the background. The choice is made by classifying the scene type.
Model accuracy:
Model achieves 91.3% accuracy on the validation set.
Classes info
The output can be the following types:
- hard
- soft
- digital*
The hard class denotes a group of scenes to which a coarser background removal method should be applied, intended for objects with an edge without small details.
The hard class contains the following categories of objects: object, laptop, charger, pc mouse, pc, rocks, table, bed, box, sneakers, ship, wire, guitar, fork, spoon, plate, keyboard, car, bus, screwdriver, ball, door, flower, clocks, fruit , food, robot.
The soft class denotes a group of scenes to which you want to apply a soft background removal method intended for people, hair, clothes, and other similar types of objects.
The soft class contains the following categories of objects: animal, people, human, man, woman, t-shirt, hairs, hair, dog, cat, monkey, cow, medusa, clothes
The digital* class denotes a group of images with digital graphics, such as screenshots, logos, and so on.
The digital class contains the following categories of scenes: screenshot
* - Not fully supported at this moment
More details on model here: https://huggingface.co/Carve/scene_classifier/
Architecture
The classifier uses DenseNet161 as the encoder and some linear layers at classifier base.
Some examples
