CrossRectify-SSOD

0. Introduction

Official code of "CrossRectify: Leveraging Disagreement for Semi-supervised Object Detection".

This repo includes training SSD300 on Pascal VOC, training Faster-RCNN-FPN on Pascal VOC, and training Faster-RCNN-FPN on MS-COCO.

The scripts about training SSD300 are based on ssd.pytorch, contained in SSD300.

The scripts about training Faster-RCNN-FPN on Pascal VOC are based on the official Detectron2 (v0.4) repo, contained in detectron2.

The scripts about training Faster-RCNN-FPN on MS-COCO are based on the official MMDetection (v2.17.0) repo, contained in mmdetection.

1. Environment

Python = 3.6.8

CUDA Version = 10.1

Pytorch Version = 1.6.0

detectron2 = 0.4 (training Faster-RCNN-FPN on Pascal VOC)

mmdetection = 2.17.0 (training Faster-RCNN-FPN on MS-COCO)

2. Prepare Dataset

Download and extract the Pascal VOC dataset.

For training SSD300 on Pascal VOC, go into the SSD300 subdirectory and specify the VOC_ROOT variable in data/voc0712.py and data/voc07_consistency.py as /path/to/dataset/VOCdevkit/

For training Faster-RCNN-FPN on Pascal VOC, go into the detectron2 subdirectory and set the environmental variable in this way: export DETECTRON2_DATASETS=/path/to/dataset/VOCdevkit/

Download and extract the MS-COCO dataset.

For training Faster-RCNN-FPN on MS-COCO, go into the mmdetection subdirectory and follow the instructions here.

3. Instructions

3.1 Reproduce Table. 1

Go into the SSD300 subdirectory, then run the following scripts to train detectors.

• fully-supervised training (VOC 07 labeled, without extra augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_ssd.py --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, without extra augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo39.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• fully-supervised training (VOC 0712 labeled, without extra augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_ssd0712.py --save_interval 12000

• fully-supervised training (VOC 07 labeled, with horizontal flip): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_csd_sup2.py --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, with horizontal flip): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_csd.py --save_interval 12000

• fully-supervised training (VOC 0712 labeled, with horizontal flip): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_csd_sup_0712.py --save_interval 12000

• fully-supervised training (VOC 07 labeled, with mix-up augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_isd_sup2.py --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, with mix-up augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_only_isd.py --save_interval 12000

• fully-supervised training (VOC 0712 labeled, with mix-up augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_isd_sup_0712.py --save_interval 12000

To eval the trained SSD300 on the Pascal VOC test set, run the following script:

• CUDA_VISIBLE_DEVICES=0 python3 eval.py --trained_model /path/to/trained/detector/ckpt.pth

3.2 Reproduce Table. 2

Go into the SSD300 subdirectory, then run the following scripts.

• fully-supervised training (VOC 07 labeled, without extra augmentation): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_ssd.py --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold = 0.5): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo39.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold = 0.8): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo39-0.8.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold increasing from 0.5 to 0.8): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo39-0.5-to-0.8.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.5, use TP and discard FP): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo36.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.5, use TP and random labeled FP, confidence threshold=0.5): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo102.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

• self-labeling (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.5, use GT labels for TP and FP, confidence threshold=0.5): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo32.py --resume weights/ssd300_12000.pth --ramp --save_interval 12000

To eval the trained SSD300 on the Pascal VOC test set, run the following script:

• CUDA_VISIBLE_DEVICES=0 python3 eval.py --trained_model /path/to/trained/detector/ckpt.pth

3.3 Reproduce Table.3

Go into the SSD300 subdirectory, then run the following scripts.

• CrossRectify (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.5): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo137.py --resume weights/ssd300_12000.pth --resume2 weights/default/ssd300_12000.2.pth --save_interval 12000 --ramp --ema_rate 0.99 --ema_step 10

• CrossRectify + mix-up augmentation (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.5): CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 train_pseudo151.py --resume weights/ssd300_12000.pth --resume2 weights/default/ssd300_12000.2.pth --save_interval 12000 --ramp --ema_rate 0.99 --ema_step 10

To eval the trained SSD300 on the Pascal VOC test set, run the following script:

• CUDA_VISIBLE_DEVICES=0 python3 eval.py --trained_model /path/to/trained/detector/ckpt.pth

Go into the detectron2 subdirectory, then run the following script.

• CrossRectify (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.7): python3 train_net.py --resume --num-gpus 8 --config configs/voc/voc07_voc12.yaml MODEL.WEIGHTS output/model_0005999.pth SOLVER.CHECKPOINT_PERIOD 36000

To eval the trained Faster-RCNN-FPN on the Pascal VOC test set, run the following script:

• python3 train_net.py --eval-only --num-gpus 8 --config configs/voc/voc07_voc12.yaml MODEL.WEIGHTS /path/to/trained/detector/ckpt.pth

3.4 Reproduce Table.4

Go into the mmdetection subdirectory, then run the following script.

• CrossRectify (VOC 07 labeled + VOC 12 unlabeled, confidence threshold=0.9): bash tools/dist_train_partially.sh semi 1 10 8

To eval the trained Faster-RCNN-FPN on the MS-COCO test set, run the following script:

• bash tools/dist_test.sh work_dirs/soft_teacher_faster_rcnn_r50_caffe_fpn_coco_180k/10/1/soft_teacher_faster_rcnn_r50_caffe_fpn_coco_180k.py work_dirs/soft_teacher_faster_rcnn_r50_caffe_fpn_coco_180k/10/1/iter_180000.pth 8 --eval bbox

Citation

If you find this work useful, please consider citing our paper. We provide a BibTeX entry of our paper below:

@article{CrossRectify,
  title={CrossRectify: Leveraging disagreement for semi-supervised object detection},
  author={Ma, Chengcheng and Pan, Xingjia and Ye, Qixiang and Tang, Fan and Dong, Weiming and Xu, Changsheng},
  journal={Pattern Recognition},
  volume={137},
  pages={109280},
  year={2023},
  publisher={Elsevier}
}

Contact

Feel free to contact [machengcheng2016@gmail.com] if you have any questions about our paper or codes.