README.md

MEPU-OWOD

Code Implementation of "Unsupervised Recognition of Unknown Objects for Open-World Object Detection" (arXiv 2308.16527)

Introduction

Our Key Technique: Reconstruction Error-based Weibull modeling (REW)

Dataset Preparation

Please first download the MS COCO dataset and the directory structure should be like:

mepu-owod/
└── datasets/
    └── coco/
        ├── annotations/
        ├── train2017/
        └── val2017/

Prepare dataset for S-OWOD:

sh prepare_dataset.sh

The training dataset for S-OWODB and M-OWODB should be like:

mepu-owod/
└── datasets/
    └── mowod/
        ├── Annotations/
        ├── ImageSets/
        └── JPEGImages/
    └── sowod/
        ├── Annotations/
        ├── ImageSets/
        └── JPEGImages/

Training and Testing

Unsupervised pretraining of REW:

sh train_rew.sh

Train MEPU on the S-OWOD benchmark:

sh script/train_mepu_fs.sh

Evaluate MEPU on the S-OWOD benchmark:

sh script/eval_owod.sh

Results

	Task1		Task2		Task3		Task4
Method	U-Recall	mAP	U-Recall	mAP	U-Recall	mAP	mAP
ORE-EBUI	1.5	61.4	3.9	40.6	3.6	33.7	31.8
OW-DETR	5.7	71.5	6.2	43.8	6.9	38.5	33.1
PROB	17.6	73.5	22.3	50.4	24.8	42.0	39.9
CAT	24.0	74.2	23.0	50.7	24.6	45.0	42.8
MEPU-FS (Ours)	37.9	74.3	35.8	54.3	35.7	46.2	41.2
MEPU-SS (Ours)	33.3	74.2	34.2	53.6	33.6	45.8	40.8