This project aims to classify accessibility feature of storefront on sidewalk from google street view images and it can help visually impaired people to avoid dangerous obstacles in the street and allow them to access each store. We are conducting some experiments on Faster-RCNN, a popular architecture in Object Detection. It enables us to have better understanding of how this model actually “sees” a physical object.
- Background(0)
- Door(1)
- Knob(2)
- Stairs(3)
- Ramp(4)
- Training Set: 928 images with labels
- Validation Set: 100 images with labels
*************************Recall Precision **************************************
Door -> TP: 159 Predict: 350 Truth: 164 Precision:45.43% Recall:96.95%
Knob -> TP: 58 Predict: 211 Truth: 77 Precision:27.49% Recall:75.32%
Stairs -> TP: 94 Predict: 361 Truth: 96 Precision:26.04% Recall:97.92%
Ramp -> TP: 1 Predict: 13 Truth: 10 Precision:7.69% Recall:10.00%
*******************************************************************************
*************************Recall Precision **************************************
Door -> TP: 154 Predict: 283 Truth: 164 Precision:54.42% Recall:93.90%
Knob -> TP: 58 Predict: 211 Truth: 77 Precision:27.49% Recall:75.32%
Stairs -> TP: 94 Predict: 361 Truth: 96 Precision:26.04% Recall:97.92%
Ramp -> TP: 1 Predict: 13 Truth: 10 Precision:7.69% Recall:10.00%
| Non-filtering | Depth-Filtering |
|---|---|
 |
 |
