CI-SSOD

The official code for our ICCV2023 paper "Gradient-based Sampling for Class Imbalanced Semi-supervised Object Detection".

Usage:

Date preparation

MS-COCO sub-task

Download the COCO2017 dataset(including training and val images) from this link.
Download the annotations for split1 and split2 for from Google Drive.
Organize the data as the following structure(or rewrite the path in configs as you need):

CISSOD/
    dataset/
        coco/
            train2017/
            val2017/
            annotations/
                instances_train2017_coco_split1_label.json
                instances_train2017_coco_split1_unlabel.json
                instances_train2017_coco_split2_label.json
                instances_train2017_coco_split2_unlabel.json

MS-COCO → Object365

Download the COCO2017 dataset(including training and instances_train2017.json) from this link.
Download the Object365dataset(including training and val images) from .this link
Download the annotations from Google Drive.
Organize the data as the following structure(or rewrite the path in configs as you need):

CISSOD/
    dataset/
        coco/
            train2017/
            val2017/
            annotations/
                instances_train2017.json

        Object365/
            images/
            annotations/
                zhiyuan_objv2_train_object365_label.json
                zhiyuan_objv2_train_object365_unlabel.json
                zhiyuan_objv2_val.json

LVIS

Download the COCO2017 dataset(including training and val images) from this link.
Download the LVIS annotations from .
Download the annotations from Google Drive.
Organize the data as the following structure(or rewrite the path in configs as you need):

CISSOD/
    dataset/
        coco/
            train2017/
            val2017/
        lvis/
            annotations/
                lvis_v1_train_lvis_label.json
                lvis_v1_train_lvis_unlabel.json
                lvis_v1_val.json

Installation:

You can follow the Soft teacher to finish the installation. Note that we do not use the wandb. Or use the following command:

make install

Note that you may need to modify the version of torch, torchvision and mmcv-full based on your cuda version. mmcv-full can be installed by:

pip install mmcv-full==1.3.12 -f https://download.openmmlab.com/mmcv/dist/YOUR_CUDA_VERSION/YOUR_TORCH_VERSION/index.html
# mmcv > 1.3.8 and <1.4.0
# pip install mmcv-full==1.3.12  -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.9.0/index.html

Training:

sh ${script} baseline/semi ${name_of_fold} ${name_of_data} ${gpu_number}

For example, for the split 1 of COCO on 8 GPUs on semi-supervised setting,

sh tools/dist_train_partially.sh semi grs coco1 8

To modify the split of COCO, change the SPLIT in line 7 of ./configs/soft_teacher/ubsoft_teacher_grs_faster_rcnn_r50_caffe_fpn_coco_180k.py. For concrete instructions of training on each setting, please refer to train.sh.

Evaluation:

sh tools/dist_test.sh ${name_of_config} ${location_of_checkpoint} ${gpu_number} --eval bbox

For example,

sh tools/dist_test.sh configs/soft_teacher/ubsoft_teacher_grs_faster_rcnn_r50_caffe_fpn_coco_180k.py work_dirs/ubsoft_teacher_grs_faster_rcnn_r50_caffe_fpn_coco_180k/coco2/grs2/iter_180000.pth 4 --eval bbox

Acknowledgment:

The code is heavily borrowed from Soft teacher and thanks for their contribution.

lijm48/CI-SSOD