1. Prepare dataset
   1. Raw *.tar from Supervisely
   2. Extraction
   3. Exploratory analysis
   4. Darknet-format dataset
2. Anchors calculation
3. Stage 1: YOLOv3 prototype training
4. Soft-NMS YOLOv3
5. Stage 2: Differentiation/classification
   1. Traditional approach
      1. Hand segmentation
      2. Hand identification
   2. Training: Classification
6. Grow dataset
7. Thorough training of the models
   1. Hard examples mining
   2. Models
      1. Stage 1 of two-stage model
      2. Single-shot YOLOv3
      3. Stage 2 (NN-based) of two-stage model
8. Evaluation on test set
9. Thesis draft
10. Optimisation
11. Write-up
12. N.B. Cells with boundary have detailed flowcharts in https://github.com/JinhangZhu/project-diary/tree/master/images