Visual One
Case Studies
Below, you can see some case studies we have done for computer vision applications in various fields based on some publicly available datasets from Kaggle or recently published research papers.
You can install our Python package and run your own exmples in 5 min.
You can also use this IPython notebook to train and evaluate the results for any of the above exmples or with your own data.
Pneumonia Detection
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 82%
Number of samples used for training: 5,216
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 5 | 5 | 71% |
| 25 | 25 | 81% |
| 50 | 50 | 87% |
| 100 | 100 | 85% |
Summary:
Best accuracy achieved with much fewer samples, 50 samples vs. 5000.
Significantly higher accuracy achieved (85% vs. 82%) with 200 training samples.
Pneumonia
Normal
Plant Bacterial Disease Detection
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 96%
Number of samples used for training: 2,475
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 86% |
| 5 | 5 | 86% |
| 15 | 15 | 94% |
| 25 | 25 | 100% |
Summary:
Higher accuracy achieved (100% vs. 96%) with significantly lower number of training samples, 50 samples vs. ~2500 samples.
Bacterial Disease
Healthy
Malaria Infection Detection
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 96%
Number of samples used for training: 22,049
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 88% |
| 5 | 5 | 86% |
| 15 | 15 | 88% |
| 25 | 25 | 96% |
| 50 | 50 | 96% |
Summary:
Same accuracy achieved with significantly lower number of training samples, 50 samples vs. ~22,000 samples!
Parasitized
Uninfected
Fraud Detection
Dataset from Kaggle.
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 86% |
| 5 | 5 | 88% |
| 15 | 15 | 94% |
| 25 | 25 | 92% |
| 50 | 50 | 96% |
Altered
Real
Fruit Classification
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 98%
Number of samples used for training: 10,900
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 81% |
| 5 | 5 | 85% |
| 15 | 15 | 98% |
| 25 | 25 | 98% |
| 50 | 50 | 99% |
Summary:
Same accuracy achieved using much fewer samples (30 vs 10,900.)
Good accuracy achieved using only 10 samples.
Rotten
Fresh
COVID Detection
Dataset from Github.
The best result achieved by UC-Berkeley researchers publishe here:
Accuracy: 89%
Number of samples used for training: 700
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 10 | 10 | 65% |
| 50 | 50 | 76% |
| 100 | 100 | 83% |
| 350 | 350 | 90% |
Summary:
Slightly better accuracy achieved using same number of samples.
Good accuracy achieved using 200 samples.
COVID
No COVID
Brain Tumor Detection
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 88%
Number of samples used for training: 193
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 5 | 5 | 76% |
| 15 | 15 | 86% |
| 25 | 25 | 87% |
| 50 | 50 | 90% |
Summary:
Best accuracy achieved with fewer samples, ~50 samples vs. 193.
Higher accuracy achieved (90% vs. 88%) with 100 training samples.
Tumor
Normal
Concrete Crack Detection
You can download the dataset from here.
The best result achieved by theresearchers publishe here:
F1 Score: 90%
Number of samples used for training: 400
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 99% |
| 5 | 5 | 99% |
| 15 | 15 | 100% |
| 25 | 25 | 100% |
Summary:
99% accuracy achieved using only 6 samples. 100% accuracy achieved using only 30 samples.
Cracked
Not Cracked
Garbage Classification
Dataset from Kaggle.
Multi-class; number of classes: 6
Best result achieved on Kaggle as reported here:
Accuracy: 95%
Number of samples used for training: 400/class
VEDX accuracy for different number of samples used for training:
| n_samples | Accuracy |
|---|---|
| 10/class | 80% |
| 25/class | 87% |
| 50/class | 89% |
| 100/class | 90% |
| 400/class | 96% |
Summary:
Slightly better accuracy achieved using same number of samples.
Good accuracy achieved using only 10/class samples.
Cardboard
Glass
Paper
Metal
Plastic
Trash
Land Scene Recognition
Dataset from Kaggle.
Multi-class; number of classes: 21
Best result achieved on Kaggle as reported here:
Accuracy: 97%
Number of samples used for training: 500/class
VEDX accuracy for different number of samples used for training:
| n_samples | Accuracy |
|---|---|
| 10/class | 84% |
| 25/class | 90% |
| 50/class | 93% |
| 100/class | 95% |
| 500/class | 97% |
Summary:
Best accuracy achieved using the same number of samples. Good accuracy achieved with only 10/class samples.
Agricultural Land
Freeway
Beach
Dense Residential
River
Golf Course
Medium Residential
Airport
Harbor
Indoor Scene Recognition
Dataset from MIT.
Multi-class; number of classes: 67
Best result they achieved as reported in their paper.:
Accuracy: 73%
Number of samples used for training: 400/class
* Note that the paper was published before the rise of Deep Learning.
VEDX accuracy for different number of samples used for training:
| n_samples | Accuracy |
|---|---|
| 10/class | 85% |
| 25/class | 90% |
| 50/class | 91% |
| 100/class | 92% |
| 400/class | 94% |
Summary:
Significantly higher accuracy achieved (85% vs 73%) with much fewer samples, 10/class vs. 400/class.
Subway
Bus
Shoe Store
Auditorium
Theater
Library
Lobby
Corridor
Pantry
