CIOD: Class Incremental Object Detector

Introduction

Based on Faster R-CNN, CIOD can add new class to a well-trained object detector, only use data of new classes.

Preparation

First of all, clone the code

git clone https://github.com/jinyu121/CIOD

Then, create a folder:

cd CIOD && mkdir data

And setup the environment:

• Python3
• PyTorch
• CUDA 8.0 or higher

Install all the python dependencies using conda:

conda env create -f environment.yml --name ciod-faster

Compile the cuda dependencies using following simple commands:

cd lib && ./make.sh

Train

Split the dataset

We split the classes of the dataset into class groups. For example, Pascal VOC dataset have 20 classes, and we split the classes into 4 class groups equally.

Then, we select pictures for each group in the whole dataset, forming image groups. Each image group will only contains pictures who only contains objects in the corresponding class group.

We can do it using select_pic.py.

Train

python3 ./trainval_net.py --session 10000

Test

python3 ./test_net.py --session 10000 --no_repr

Citation

@InProceedings{ICME2019:FCIOD,
    author = {Yu Hao and Yanwei Fu and Yu-Gang, Jiang and Qi, Tian,
    title = {An End-to-End Architecture for Class-Incremental Object Detection With Knowledge Distillation},
    booktitle = {IEEE International Conference on Multimedia and Expo (ICME '19)},
    year = {2019}
}

@InProceedings{ICMR2019:TGFS, 
    author = {Yu, Hao and Yanwei, Fu and Yu-Gang, Jiang}, 
    title = {Take Goods from Shelves: A Dataset for Class-Incremental Object Detection},
    booktitle = {ACM International Conference on Multimedia Retrieval (ICMR '19)},
    year = {2019}
}

@inproceedings{renNIPS15fasterrcnn,
    Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
    Title = {Faster {R-CNN}: Towards Real-Time Object Detection with Region Proposal Networks},
    Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
    Year = {2015}
}

Note

It's toooo busy to re-write the code using Detectron or pure TorchVision, which have been planned for long.

And you can also check out d58255605e37e362ee746637fe93efb669a7e7ab. Most of the results are produced based on this commit.