Our proposed systematic training and evaluation methodology provides the opportunity to examine a wide range of aspects in detail.
The necessary models and results emerge from the overall approach and can be broken down into five subareas:
- Evaluation of Image Changes
- Influence of Model Size and Data Type on Training and Evaluation
- Influence of Object Size and Frequency within the Anonymized Class
- Evaluation of Different Fine-Tuning Strategies
- Influence of Label Error
- Using SSIM
- Results are available for relevant bounding boxes and whole images of specific classes
- All model sizes trained and evaluated on both data types
- Evaluation results include mAP or AP of specific classes
- Using model size m
- Evaluation of defined objects, depending on their size and occurring frequency within the person class
- Also allows comparison with other work using specific model sizes
- Using base model of size m trained on original data as the base for fine-tuning on anonymized data
- Evaluation of differences between using labels from original data and labels modified to fit anonymized data
