Data Preparation

June 24, 2022 · View on GitHub

1. SSOD: Semi-Supervised Object Detection

We support 5 popular settings in SSOD research as listed below：

Labeled Data	Unlabeled Data	Test Data
COCO2017-train-1%	COCO2017-train-99%	COCO2017-test
COCO2017-train-5%	COCO2017-train-95%	COCO2017-test
COCO2017-train-10%	COCO2017-train-90%	COCO2017-test
COCO2017-train	COCO2017-unlabeled	COCO2017-test
VOC07-trainval	VOC12-trainval	VOC07-test

Download VOC and COCO from the website and organize them as follows:

# ====coco====                      |          # ====voc====
/data/coco/                         |          /data/voc/
  - images                          |            - 12
    - train2017                     |     		    - VOCdevkit
    - unlabeled2017                 |                - VOC2012
	- ...                           |                  - ...      
  - annotations                     |            - 07
	- instances_train2017.json      |              - VOCdevkit
	- image_info_unlabeled2017.json |                - VOC2007
	- ...						    |                  - ...

Run scripts to create the soft symlink：

# * please change the "prefix_coco", "prefix_coco_ul", "prefix_voc" in the scripts to fit your environment.
# * you can also create symlink by yourself.
cd tools/datasets
xonsh create_dataset_link.sh

Create coco-standard, coco-additional, voc (it will cost several minutes):
```
cd tools/datasets
xonsh preprocess_dataset.sh
```

2. DAOD: Domain Adaptive Object Detection

We support 4 popular settings in DAOD research as listed below:

	Labeled Data	Unlabeled Data	Test Data
normal $\to$ foggy (C2F)	cityscapes (train)	cityscapes-foggy (train)	cityscapes-foggy (val)
small $\to$ large (C2B)	cityscapes (train)	BDD100K (train)	BDD100K (val)
across cameras (K2C)	KITTI (train)	cityscapes (train)	cityscapes (val)
synthetic $\to$ real (S2C)	Sim10K	cityscapes (train)	cityscapes (val)

Download cityscapes, cityscapes-foggy, KITTI, Sim10K and BDD100K from the website and organize them as follows:

# cityscapes          |    # cityscapes-foggy      |   # BDD
/data/city            |    /data/foggycity         |   /data/BDD
  - VOC2007_citytrain |      - VOC2007_foggytrain  |     - VOC2007_bddtrain
    - ImageSets       |        - ImageSets         |       - ImageSets
    - JPEGImages      |        - JPEGImages        |       - JPEGImages
    - Annotations     |        - Annotations       |       - Annotations 
  - VOC2007_cityval   |      - VOC2007_foggyval    |     - VOC2007_bddval 
    - ImageSets       |        - ImageSets         |       - ImageSets
    - JPEGImages      |        - JPEGImages        |       - JPEGImages
    - Annotations     |        - Annotations       |       - Annotations 
# =========================================================================
# KITTI               |   # Sim10K
/data/kitti           |   /data/sim
   - ImageSets        |     - ImageSets
   - JPEGImages       |     - JPEGImages
   - Annotations      |     - Annotations

PS: please refer to ProbabilisticTeacher for the detailed dataset pre-processing.

Run scripts to create the soft symlink：

cd tools/datasets_uda
xonsh create_dataset_link.sh

Convert to coco format:

cd tools/datasets_uda
xonsh preprocess_dataset.sh