# Tasks: [09-04 to 09-11]

## Custom Object Detector:
#### Algorithm:
SSD7
#### Loss:
Multi-task log loss for classification and smooth L1 loss for localization.
#### Output classes:
Adult, Kid

#### Data Selection and Splits:
From each camera of video series - 6000 people are selected from consecutive frames.

| Video-series | Selected People |
| -------- | -------- | -------- |
| totmate_20200811 | 12,000|
| totmate_20200812 | 18,000 |
| totmate_20200817 | 12,000 |
| totmate_20200818 | 18,000 |
| totmate_20200819 | 18,000 |
| totmate_20200820 | 12,000 |
| totmate_20200821 | 18,000 |

Metadata files generated by YoloV4 and kid and adult classifier are used as GT for new detector without manual intervention for correcting errors [missing bounding boxes and erroneous bounding boxes].

Training split - 79,760 people
Validation split - 34,212 people
Total frames - 58,558

#### Data preprocessing
- All image data is resized to h=1080, w=1920 to h=300, w=480 frames
- All bounding boxes are also scaled down accordingly.

#### Data Augmentation (online- during training)
- Color based: random-brightness, contrast, hue, saturation
- Spatial: flip, translate, scale

#### Training
50,000 Iterations: 50 Epochs * 1000 steps

These are some sample visuzlisation of SSD7.
[Green bounding box - YOLOv4 predicted GT]
[Red bounding box - SSD7 predicted Kid]
[Turquoise bounding box - SSD7 predicted Adult]

Result-1:

Result-2:

Result-3:

Result-4:

Result-5:

Result-6:

Result-7:

Result-8:

#### Comments:
- Took 14 hrs of training time.
- Classification is good, localisation needs improvement
- It cannot outperform big models like YoloV4 or SSD512 in terms of accuracy, but can give decent results.

#### Inference Speed:
##### Old model- YoloV4 + Kid and Adult classifier(InceptionV3)
- YoloV4 takes 50 ms per frame ()
- Kid and Adult classifier takes 40ms per cropped image i.e, if two people are there in a frame it would take 2 * 40ms = 80ms
- Total time to run a frame with 2 people in it: 50ms + 80ms = 130ms
- Processing speed is 7 FPS

##### New model - SSD7
- SSD7 takes 23 ms per frame
- Processing speed is 43 FPS

#### H/W:
- CPU: i5-9400F
- GPU: RTX-2070Super
- RAM: 16 GB

##### Scope for improvement:
- Increase training data
- Manual Correct Algorithm predicted GT labels