OneTeacher

Official implementation of "Towards End-to-end Semi-supervised Learning for One-stage Object Detection". OneTeacher is a semi-supervised framework for YOLOV5, which is equiped with two novel designs, namely Multi-view Pseudo-label Refinement and Decoupled Semi-supervised Optimization.

Updates

• (2023/1/29) Release our project.

Data Preparation

1. Prepare the datasets of COCO and VOC, please refer to YOLOv5

After that, the file structure should look like:

   |-- OneTeacher
   |-- datasets
       |-- coco
             |-- images
                  |-- train2017
                  |-- val2017
                  |-- test2017
             |-- labels
             |-- annotations
                  |-- instances_train2017.json
                  |-- instances_val2017.json

2. Generate semi-supervised data split.

   python active_sampling/generate_random_supervised_seed_yolo.py  --dataset_name 'coco_2017_train'  --random_seeds 0,1,2,3,4,5,6,7,8,9  --random_file  ./data_processing/COCO_supervision.txt   --random_percent 10.0  --output_file ./dataseed/COCO_supervision_10.json

   Training and Evaluation

   Semi-supervised Learning on COCO 10%

   bash script/coco_semi_script.sh 32 2 0,1

   Semi-supervised Learning on VOC 25%

   bash script/voc_semi_script.sh 32 2 0,1

   Fully-supervised Learning on COCO 10%

   bash script/coco_fully_script.sh 32 2 0,1

   Fully-supervised Learning on VOC 25%

   bash script/voc_fully_script.sh 32 2 0,1