Documentation

COCO Dataset Visualization for Object Detection

Feature Space Visualization

This tool is used to put the feature vector into a two-dimensional coordinate system, and draw a feature distribution diagram for each annotation.

Script Execution Process

You can run the cut_images_by_bbox.py script to analyze the dataset. The following describes how to use the parameters.
Table 1 Image cropping parameters

Parameter

Type

Value Range

Default Value

Description

test_coco_root

String

-

./split_dataset_coco_feature/test

COCO test dataset.

train_coco_root

String

-

./split_dataset_coco_feature/train

COCO training dataset.

type_label

String

-

screw_1,screw_2, screw_3, screw_4, screw_5, mem

Label for image cropping.

class_num

Integer

> 0

6

Classification quantity of the ResNet-50 model.

pretrained_ckpt_path

String

-

../../../pre_trained_ckpt

Path for storing the pre-trained ResNet-50 model.

The command reference of the script is as follows:

python3 com_package/object_detection/data_analysis/cut_images_by_bbox.py --test_coco_root 'com_package/object_detection/data_analysis/split_dataset_coco_feature/test' --train_coco_root 'com_package/object_detection/data_analysis/split_dataset_coco_feature/train' --type_label 'screw_1'

The reference log information is as follows:

Figure 1 Log information generated during feature space visualization

The feature space visualization results are saved as JPG images, as shown in Figure 2.

Figure 2 Feature space visualization results