# Incremental learning in semantic segmentation from image labels

Incremental Learning approaces update the model's parameters by training only on new date and employing as-hoc techiniques to avoid catastrphic forgetting on old classes.

### Problems:
- struggle to update their models incremntally as new categories comes in
- pixel-by-pixel annotations are expensive and time-consuming
- existing model are not meant to be incrementally updated over time and their knowledge is often limited on the predefined set of classes

### Objective:
- incrementally updating the model
- reducing the annotation cost of new data for semantic segmentation

### Task: Weakly supervised incremental learning for semantic segmentation (WILSS)

### **W**eakly **I**ncremental **L**earning framework for semantic **S**egmentation that incrementally trains a segmentation model generating **ON**line pseudo-supervision from image-level annotations (WILSON)


**Problem definition**:
Given an input image space $X$, each image is composed by a set of pixels $I$ with constant cardinality $|I| = H * W = N$. The output Space $Y^N$ is defined as the product set of N-tuples with elements in a label space $Y$.

We want: given an image $x \in X$ learn the mapping to assign each pixel $x_i$ a label $y_i \in Y$

Output: (segmentation mask)
$$\{arg\ max_{c \in Y}\ p^c_i\}^N_{i=1} $$ where $p^c_i$ is the model prediction of pixel $i$ for class $c$.

### Training the localizer
features: from the segmentstion encoder $e$ to predict a score for all classes (background , old and new ones)

#### Learn from image-level labels
**normalized Global Weighted pooling**: weights every pixel based on its relevance for the target class