Offline Data Augmentation for a COCO Object Detection Dataset

The augmentation tool path is ./dataset_tools/coco_data_aug/coco_data_aug.py. The data augmentation for object detection supports offline data augmentation, which is implemented by mobility and blocking augmentation.

Table 1 describes the names, types, value ranges, default values, and descriptions of each parameter in coco_data_aug.

**Table 1** Parameter description
Parameter	Type	Value Range	Default Value	Description
--input_dataset_path	String	-	""	Path of the COCO dataset.
--output_dataset_path	String	-	./output_dataset_path	Path for saving the augmented data.
--x_offset	Bool	True or False	False	Mobility augmentation switch on the X axis.
--y_offset	Bool	True or False	False	Mobility augmentation switch on the Y axis.
--offset_target_times	Integer	[1,10]	1	Increased multiplier of data for mobility augmentation.
--obj_block	Bool	True or False	False	Blocking augmentation switch.
--block_labels	String	-	label0,label1	List of blocked object labels. Use commas (,) to separate multiple labels.
--obj_erase	Bool	True or False	False	Erasing switch for blocking augmentation.

The following is an example of the default parameter settings:

python3 ./dataset_tools/coco_data_aug/coco_data_aug.py --input_dataset_path=./train --output_dataset_path=./output_dataset_path --x_offset=True --y_offset=True --obj_block=True --block_labels='label0,label1' --obj_erase=True --offset_target_times=2

Each augmentation mode can be enabled separately or in combination. If the blocking augmentation is enabled, you need to specify the tag to be blocked in --block_labels.

The reference log information is as follows:

Figure 1 Log information

After the script is executed, augmented data is generated in the output directory specified by the --output_dataset_path parameter.

Figure 2 Specified directory

Parent topic: Dataset Processing Tool Package