Skip to content

Latest commit

 

History

History
101 lines (63 loc) · 4.86 KB

README.md

File metadata and controls

101 lines (63 loc) · 4.86 KB
Raw

Classes Stats for Images

Overview • How To Run • Explanation

GitHub release (latest SemVer) views runs

Overview

Data Exploration for Segmentation and Detection tasks is underestimated by many researchers. The accuracy of your models highly depends on how good you understand data.

This app "Classes Stats for Images" generates report with detailed general and per image statistics for all classes in images project. It allows to see big picture as well as shed light on hidden patterns and edge cases (see How to use section).

How To Run

Step 1: Run from context menu of project / dataset

Go to "Context Menu" (images project or dataset) -> "Report" -> "Classes stats for images"

Step 2: Configure running settings

Choose the percentage of images that should be randomly sampled. By default all images will be used. And then press "Run" button. In advanced settings you can change agent that will host the app and change version (latest available version is used by default).

Step 3: Open app

Once app is started, new task appear in workspace tasks. Monitor progress from both "Tasks" list and from application page. To open report in a new tab click "Open" button.

App saves resulting report to "Files": /reports/classes_stats/{USER_LOGIN}/{WORKSPACE_NAME}/{PROJECT_NAME}.lnk. To open report file in future use "Right mouse click" -> "Open".

Explanation

Per Image Stats

Columns:

  • IMAGE ID - image id in Supervisely Instance
  • IMAGE - image name with direct link to annotation tool. You can use table to find some anomalies or edge cases in your data by sorting different columns and then quickly open images with annotations to investigate deeper.
  • HEIGHT, WIDTH - image resolution in pixels
  • CHANNELS - number of image channels
  • UNLABELED - percentage of pixels (image area)

Columns for every class:

  • - class area (%)
  • - number of objects of a given class (%)

Per Class Stats

  • CLASS NAME
  • IMAGES COUNT - total number of images that have at least one object of a given class
  • OBJECTS COUNT - total number of objects of a given class
  • AVG CLASS AREA PER IMAGE (%) -
              the sum of a class area on all images               
 -------------------------------------------------------------- 
 the number of images with at least one object of a given class
  • AVG OBJECTS COUNT PER IMAGE (%) -
              total number of class objects               
 -------------------------------------------------------------- 
 the number of images with at least one object of a given class

Histogram: AVG AREA / AVG OBJECTS COUNT

Histogram view for two metrics from previous chapter: AVG CLASS AREA PER IMAGE (%) and AVG OBJECTS COUNT PER IMAGE (%)

Images Count With / Without Class

TOP-10 Image Resolutions