author ：18 the cyl

Time ：2021-08-15

The paper ：《Towords Open World Object Detection》

One 、 The problem background ：

1) The test image may contain objects from unknown classes , Should be classified as unknown .
2) When information about these identified unknowns （ label ） When available , The model should be able to learn new classes step by step . And don't forget the previously learned classes . namely Open world object detection

Classes that have not been marked in the past will be explicitly trained as a background , But even the most advanced classifiers have a very high false positive rate .

Two 、 resolvent ：

There has been research on incremental model to solve this problem ,
This paper presents a new method ：ORE
It's based on contrast clustering 、 Energy based unknown identification

3、 ... and 、 Specific principle ：

ORE:
1、 Set up an object detection model MC, Trained to detect all previously encountered object classes .
2、 And you can classify new and unseen class instances as unknown .
3、 Send a large number of unknown class collections to people , Provide tags and training examples of new classes of interest .
4、 Incrementally increase... From the model n Create a new class and update yourself to produce a new model MCn, The advantage is that you can no longer train from scratch

5、 The method to make the model learn new classes without forgetting the old classes is ： Store a set of balanced examples , And adjust the base note after each increment step. .（ Wen Zhong , Using only one new class for training will reduce the accuracy of the model in identifying the original old class , At the same time, other studies have found that the effectiveness of storing a small number of instances and readjusting the model is very high ）