Documentation

SSD Small Object Detection

The ssd_tiled_dataset_mindspore folder is added for SSD small object detection based on product applications. Compared with the ssd_mobilenet_fpn_mindspore folder, only the on_platform/plat_cfg.yaml file is different. You can add extra command parameters to enable this function. The ssd_tiled_dataset_mindspore folder is recommended.

Training Parameters and Value Ranges

The following parameters are added to the SSD small object detection training parameters in Table 1. For details, see Table 1.

Table 1 Training parameters for SSD small object detection (ssd_mobilenet_fpn_mindspore/tiled_dataset_train.py)

Parameter

Type

Value Range

Default Value

Description

--split_image

Bool

True or False

True

Whether to enable small object detection.

--split_auto

Bool

True or False

False

Whether to automatically set parameters for small object detection.

--split_resize_ratio

String

Range of the two numbers in a value: [0.1,10].

"2,2"

Scaling ratio of the original image. The two numbers in the value indicate the width and height.

--split_block_size

String

Range of the two numbers in a value: [256,2048].

"896,896"

Size of image blocks. The two numbers in the value indicate the width and height.

--split_overlap_size

String

Range of the two values: [128, minimum value of split_block_size-128]

"256,256"

Size of the overlapping area of image blocks. The two numbers in the value indicate the width and height.

Training Command Reference

python3 tiled_dataset_train.py --train_dataset_path={dataset_path} --train_output_path={output_path} --pretrained_ckpt_path={path_of_the_pre-trained_model} --epoch_size=100 --batch_size=4 --input_width=2048 --input_height=1536 --init_lr=0.001 --device_num=1 --run_eval=True --eval_start_epoch=5 --split_resize_ratio=4,4 --split_block_size=896,896 --split_overlap_size=128,128 --split_auto=False --split_image=True

Evaluation Parameters and Value Ranges

The evaluation script is stored in /mxAOITraining/ssd_mobilenet_fpn_mindspore/tiled_dataset_eval.py.

Table 2 describes the evaluation parameters for SSD small object detection.

Evaluation Command Reference

python3 tiled_dataset_eval.py --eval_dataset_path={evaluation_dataset_path} --eval_ckpt_path={output_path} --eval_output_path=./eval_result