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

This directory contains the synthetic part of the SYNOSIS dataset.

## Rendering details
- Rendering using Appleseed path tracing.
- Camera: Pinhole
- Viewpoints depent on testbody orientation (B/M), but are a x-axis mirrored version of the other. Viewpoints correspond to the selected viewpoints in the real dataset.
  * 3 viewpoints from the perpendicular, in one direction with 10 degree separation (0, 10, 20).
- SPP: 256
- Bounces (total): 8
- Light samples: 4

### Image files
Rendered images and masks are stored grayscale in linear colorspace (no gamma).
Masks are index-based (see label details below).

### Rendered objects
- TB_correct_1 - correct, sandblasted, orientation B
- TB_correct_2 - correct, sandblasted, orientation M
- TB_correct_3 - correct, parallel, orientation B
- TB_correct_4 - correct, parallel, orientation M
- TB_correct_5 - correct, parallel, orientation B
- TB_correct_6 - correct, parallel, orientation M
- TB_correct_7 - correct, spiral, orientation B
- TB_correct_8 - correct, spiral, orientation M
- TB_correct_9 - correct, spiral, orientation B
- TB_correct_10 - correct, spiral, orientation M
- TB_defect_1 - defect, sandblasted, orientation B
- TB_defect_2 - defect, sandblasted, orientation M
- TB_defect_3 - defect, parallel, orientation B
- TB_defect_4 - defect, parallel, orientation M
- TB_defect_5 - defect, parallel, orientation B
- TB_defect_6 - defect, parallel, orientation M
- TB_defect_7 - defect, spiral, orientation B
- TB_defect_8 - defect, spiral, orientation M
- TB_defect_9 - defect, spiral, orientation B
- TB_defect_10 - defect, spiral, orientation M

## Dataset Structure
```
labels
    |-- B
      |-- segmentation_masks
      |-- instance_masks
    |-- M
      |-- segmentation_masks
      |-- instance_masks
TB_*_*
    |-- configs
      |-- defect_classes.csv
      |-- defect_instances.csv
    |-- images
        |-- instance_name_0
          |-- render_000.png
          |-- ...
        |-- ...
...
```

### Descriptions
- `TB_*_*/images` - Rendered images for each of the generated instances. Correct geometries differ in instances only through the choice of texture instances of each of the faces.
- `labels` - The labels are the same across all defective test bodies and their geometry instances. Thus they are stored separately to reduce memory footprint.

### Statistics of annotated subset
- Defect classes:

  0. background
  1. bigdent
  2. complexscratch
  3. flatscratch
  4. smalldent

- Number of images:
  - all: 7200
  - sandblasted: 1440
  - parallel: 2880
  - spiral: 2880

### Label details
Labels are tied to geometric instances and are share among all testbodies of the same orientation (B/M).

Generated segmentation masks are stored in uint8 index-based format, meaning the pixel value corresponds to the class/instance index in the statistics above.
Masks were postprocesses to remove annotations with contrast lower than 0.05, where we define the contrast as the difference between the mean intansity of the defect and its surrounding ring (2 times dilated mask subtracting original mask).
This approximately matches the annotations of human annotators.