Awesome
SIOD: Single Instance Annotated Per Category Per Image for Object Detection
Main Results
| Detector | Task | $AP@\mathbb{S}$ | $AP@\mathbb{S}_0$ | $AP@\mathbb{S}_3$ | $AP@\mathbb{S}_5$ | $AP@\mathbb{S}_7$ | $AP@\mathbb{S}_9$ |
|---|---|---|---|---|---|---|---|
| CenterNet-Res18 | FSOD | 17.3 | 28.1 | 24.0 | 17.1 | 8.8 | 1.5 |
| CenterNet-Res18 | SIOD(base) | 13.9 | 25.1 | 18.5 | 12.3 | 6.1 | 1.4 |
| CetnerNet-Res18 | SIOD(DMiner) | 16.8(+2.9) | 26.6(+1.5) | 22.4(+3.9) | 17.1(+4.8) | 9.4(+3.3) | 2.1(+0.7) |
| CenterNet-Res101 | FSOD | 22.6 | 34.2 | 30.3 | 23.6 | 13.6 | 3.1 |
| CenterNet-Res101 | SIOD(base) | 15.1 | 27.8 | 20.9 | 13.3 | 6.1 | 1.1 |
| CenterNet-Res101 | SIOD(DMiner) | 19.7(+4.6) | 29.8(+2.0) | 26.0(+5.1) | 20.5(+7.2) | 12.2(+6.1) | 2.9(+1.8) |
Dataset
-
Keep1_COCO2017_Train: keep1_instances_train2017.json
-
Semi-supervised annotation which has equivalent instance annotations to Keep1_COCO2017_Train: mark_semi_instances_train2017.json. (For Table 2)
We add a new field("keep") to the image infomation in annotation file, where keep=1 indicates the image belongs to labeled part and keep=0 indicates the image belongs to unlabeled part.
Preparations
1. pip install -r requirements.txt
2. install pytorch=1.7.0(higher version has some problems in following installation of dcnv2)
3. install dcnv2
cd src/lib/models/networks/DCNv2
sh make.sh
4. install cocoapi
cd src/lib/datasets/dataset/cocoapi/
sh install.sh
5. install nms
cd src/lib/external
make
6. create soft link for the data
vim link.sh
sh link.sh
Training
Take CenterNet-Res18 for example:
-
Directly train the centernet under SIOD setup.
sh base_resdcn18_train.sh -
Train the centernet equipped with SPLG or PGCL.
# SPLG sh plg_resdcn18_train.sh # PGCL gcl_resdcn18_train.sh -
Train the centernet equipped with DMiner.
dminer_resdcn18_train.sh or all_resdcn18_train.sh
Evaluation
Evaluate the detector with new Score-aware Detection Evaluation Protocol.
# modify the parameter "load_model" accordingly
sh test_resdcn18.sh
Visualization
Prepare some images and modified visualize.sh accordingly.
sh visualize.sh