# [09-11 to 09-18]
## Task - 1: Dataset Annotation and scraping:
- Collected ~1100 Not-appropriate images from milcube - 818 video series, checked 70k images
- Scraped 18,000 not-appropriate images online and filtered bad ones. Obtained 4500 not-app images

## Task - 2: In-Appropriate kids Detector
Rules of not-appropriate:
1. Either kid is completely naked,
2. wearing just shirt no pants,
3. wearing just pants no shirt,
4. wearing just diaper,
5. pants stretched such that under ware is visible,or shirt lifted such that tummy is visible.

Definition of Half naked: (2), (3), (5)
Definition of complete naked: (1), (4)

## Experiment - 1

#### Algorithm:
AlexNet
#### Loss:
binary crossentropy loss
### Activation:
Sigmoid
#### Output classes:
appropriate, not-appropriate

### Dataset collection and splits

Sources -
1. scraped from getty images with keys - `naked-kids`, `shirtless-kids`, `diaper-baby`
2. Milcube data

Splits -
Training: app - 5,493; not-app - 5,380;
Validation: app - 1,344; not-app - 1,366;

### Results

Accuracy on validation data: 80.87% (706/873)
Accuracy on getty images data: 91.6% (198/216)

## Experiment - 2

#### Algorithm:
AlexNet
#### Loss
Categorical crossentropy
#### Activation:
Softmax
#### Output classes:
appropriate, completely-naked, half-naked

### Dataset collection and splits

Sources -
1. scraped from getty images with keys - `naked-kids`, `shirtless-kids`, `diaper-baby`
2. Milcube data

Splits -
Training: app - 1,746; completely-naked - 1845; half-naked - 1685
Validation: app - 314; completely-naked - 205; half-naked - 354

### Results

Accuracy on validation data: 81.4% (706/873)

Total number of test images = 873
Total number of exactly correct images = 663

Exact correct images accuracy = 76%

No. of half naked predicted as completely naked - 20
No. of completely naked predicted as half naked - 28
considering half naked and completely naked as not-appropriate
app and not-app accuracy: exactly correct images(663) + not-appropriate(20 + 28) = 711

Total accuracy = 711/873 = 81.4%

## Conclusion:

- Experiment-2 seems effective we can set a strict threshold for completely-naked class since it is very sensitive.
- Data deficit - Needs more training data
- Fail cases :
- wearing Shorts
- Rule-5
- Rule-2
- Rule-5