# SYNOSIS project - Real version of the dataset
- Date: 2023-10-15
- Author: Juraj Fulir

This directory contains the annotated real part of the SYNOSIS dataset.


## Acquisition details

### Arch
- 1 arch per face:
  * 3 viewpoints from the perpendicular, in one direction with 10 degree separation (0, 10, 20).
  * All arches are aligned along the object x-axis
- Camera exposures: 150

Reference positions:
- M bodies: viewpoint idx 0
- B bodies: viewpoint idx 0

### Acquired objects
- 01B - correct, sandblasted
- 02M - correct, sandblasted
- 03B - correct, parallel
- 04M - correct, parallel
- 05B - correct, parallel
- 06M - correct, parallel
- 07B - correct, spiral
- 08M - correct, spiral
- 09B - correct, spiral
- 10M - correct, spiral
- 13B - defected, sandblasted
- 14M - defected, sandblasted
- 15B - defected, parallel
- 16M - defected, parallel
- 17B - defected, parallel
- 18M - defected, parallel
- 19B - defected, spiral
- 20M - defected, spiral
- 21B - defected, spiral
- 22M - defected, spiral
- 25B - defected, sandblasted
- 26M - defected, sandblasted
- 27B - defected, parallel
- 28M - defected, parallel
- 29B - defected, parallel
- 30M - defected, parallel
- 31B - defected, spiral
- 32M - defected, spiral
- 33B - defected, spiral
- 34M - defected, spiral

## Dataset Structure
```
150_annotated
    |-- 01B
        |-- images
            |-- Image_0_0.png
            ...
    ...
    |-- 13B
        |-- bboxes
            |-- Image_0_0.bbox
            ...
        |-- classes
            |-- Image_0_0.csv
            ...
        |-- configs
            |-- defect_classes.csv
            ...
        |-- images
            |-- Image_0_0.png
            |-- Image_0_0.json
            ...
        |-- instance_masks
            |-- Image_0_0.png
            ...
        |-- segmentation_masks
            |-- Image_0_0.png
            ...
```

### Descriptions
- `150_annotated` - Acquired images that were manually annotated using labelme. Contains the images and labels files (classes, masks, bboxes) generated from labelme jsons.

### Statistics of annotated subset
- Classes (with label indices):

    0. background
    1. dent
    2. irr_scratch
    3. scratch
    4. uncategorized

- Number of images (per texture group):
  - all: 270
  - sandblasted: 54
  - parallel: 108
  - spiral: 108

### Label details
Generated segmentation/instance masks are index-based, meaning the pixel value corresponds to the class/instance index in the `*/configs/defect_classes.csv` file.
The value `0` is implicitly reserved for the background class.
Instance indices are generated randomly per-image and don't correspond to the physical instances in the object.

Generated classes files contain the number of pixels in segmentation mask corresponding to that class.
The first value corresponds to the backgorund class.
Use `int(mask>0)` thresholding to obtain the hard (binarized) multi-label class vector.

Generated bbox files follow the format `[x_min, y_min, x_max, y_max, cls_name]` in pixels, where `cls_name` is the string name of the class (as mentioned in statistics).