Analysis results
Interactive distribution plot
Target size impact
Recommendations and warnings
Detailed image records
Machine-readable summary
Run the calculator to generate a reusable summary.
Formula used
These formulas summarize image geometry and storage behavior. They also estimate resizing impact before model training. Use results as practical dataset diagnostics.
How to use
- Add images, upload metadata, or load sample records.
- Select a plot, measurement, outlier method, and target size.
- Apply optional filters for classes, formats, splits, or dimensions.
- Press the analysis button to calculate results.
- Inspect charts, statistics, outliers, resizing needs, and warnings.
- Export the chart, data, summary, or PDF report.
Example data format
| Image | Width | Height | File size KB | Format | Class | Split |
|---|---|---|---|---|---|---|
| image_001.jpg | 640 | 480 | 125 | JPG | Cat | Train |
| image_002.png | 1024 | 768 | 840 | PNG | Dog | Train |
| image_003.jpg | 224 | 224 | 42 | JPG | Cat | Validation |
| image_004.webp | 1920 | 1080 | 310 | WebP | Dog | Test |
Understanding image-size distributions
Image dimensions directly affect memory and training speed. Large differences can create unstable preprocessing pipelines. Consistent sizing simplifies batching and deployment.
Resizing may remove details or introduce unwanted distortion. Padding preserves content while adding empty regions. Cropping preserves scale but can remove subjects.
Aspect-ratio checks reveal stretched or unusual samples. Class comparisons can expose collection and labeling bias. Split comparisons help identify distribution shift.
Frequently asked questions
Why analyze image sizes before training?
Size analysis reveals inconsistent inputs and expensive samples. It guides resizing, cropping, padding, and batching choices. Early checks prevent avoidable training failures.
What is a useful target size?
The best target depends on model architecture. Common values include 224, 256, and 512 pixels. Test accuracy and memory before final selection.
Should images always become square?
Square inputs simplify many convolutional model pipelines. They may require padding or content-aware cropping. Detection models often preserve more varied ratios.
What does an aspect-ratio outlier indicate?
It may be a panorama or narrow portrait. It can also signal corrupted metadata. Review unusual records before removing them.
When should I use logarithmic axes?
Logarithmic axes help with very wide ranges. They reveal smaller groups beside extreme values. Zero values cannot appear on log scales.
Why compare training and testing splits?
Different dimensions can create preprocessing distribution shift. The model may then face unfamiliar resolution patterns. Similar splits usually support fairer evaluation.
Does file size equal image quality?
File size depends on format and compression. Dimensions and visual detail also affect storage. Compare several indicators before judging quality.
Can folder uploads work everywhere?
Modern Chromium browsers usually support folder selection. Other browsers may only support multiple files. The standard file selector remains available.
How are duplicate records detected?
The calculator compares normalized image names and dimensions. Exact metadata duplicates receive a warning flag. Duplicate policy controls import behavior.