transform a fixed feature extractor into an Instance feature extractor that produces discriminative embeddings that are aligned with the per-class representatives
Cosine-Similarity classifier
TFA -> MTFA -> iMTFA
TFA(Two-stage Fine Tuning)
First stage, Feature Extractor: Faster RCNN(trained on the base classes)
: RoI classifier, a fully connected layer which produce classification score -> the class with lowest cosine distance (btw an RoI's embedding and the class representatives)
On iMTFA: Class-agnostic manner
can add new classes by simply averaging their computed embeddings and placing them in the classification head's weight matrix
Incremental Few-Shot Instance Segmentation
논문링크, Model Name: iMTFA (Incremental MTFA), code 지금 하고 있는 연구에서 확장하고자 다음과 같은 근거로 읽게 되었다.
Problem
Goal
In this paper
TFA(Two-stage Fine Tuning)
MTFA(Mask-TFA)
Second stage, Prediction Head
Cosine-similarity classifier
: RoI classifier, a fully connected layer which produce classification score -> the class with lowest cosine distance (btw an RoI's embedding and the class representatives)