Conceitos Básicos
Leveraging tread depth maps from online retailer images to effectively match crime-scene shoeprints and retrieve the most likely shoe models.
Resumo
The paper introduces a method called CriSp to address the problem of matching crime-scene shoeprints to a database of shoe tread depth maps. The key insights are:
- Matching crime-scene shoeprints to tread depth maps, rather than clean reference prints, leads to better retrieval performance.
- CriSp uses a data augmentation module to simulate noisy and occluded crime-scene shoeprints from clean prints in the training data.
- CriSp employs a spatial encoder to ensure the model learns to match patterns in corresponding regions of shoe treads.
- CriSp incorporates a feature masking module to focus on the visible portions of partially occluded crime-scene shoeprints during retrieval.
The authors create a large-scale reference database of tread depth maps and two validation sets of real and simulated crime-scene shoeprints. CriSp significantly outperforms state-of-the-art methods in both automated shoeprint matching and image retrieval tailored to this task.
Estatísticas
The paper reports the following key statistics:
The training dataset (train-set) contains 21,699 shoe instances from 4,932 different shoe models.
The reference database (ref-db) contains 56,847 shoe instances from 24,766 different shoe models.
The val-FID validation set contains 106 real crime-scene shoeprints linked to 1,152 shoe models and 2,770 shoe instances in ref-db.
The val-ShoeCase validation set contains 146 simulated crime-scene shoeprints linked to 16 shoe models and 52 shoe instances in ref-db.
Citações
"Shoeprints are more likely to be found at crime scenes, though they may possess fewer distinct identifying features compared to other biometric samples like blood or hair."
"Matching directly to RGB tread images causes models to overfit to irrelevant details such as albedo and lighting."
"Our data augmentation module in combination with our training set of paired tread depth maps and clean prints serves as a viable alternative to an ideal dataset containing paired crime-scene shoeprints and corresponding tread depth maps."