Segmentation network to classify each pixel into 4 classes (BG, road, curb, mark).
| Metric | Value |
|---|---|
| Image size | 896x512 |
| GFlops | 4.7 |
| MParams | 0.18 |
| Source framework | PyTorch* |
The quality metrics calculated on 500 images from "Mighty AI" dataset that was converted for 4 class classification task are:
| Label | IOU | ACC |
|---|---|---|
| mean | 84.4% | 90.1% |
| BG | 98.6% | 99.4% |
| road | 95.4% | 97.4% |
| curbs | 72.7% | 83.1% |
| marks | 70.8% | 80.6% |
IOU=TP/(TP+FN+FP)ACC=TP/GTTP- number of true positive pixels for given classFN- number of false negative pixels for given classFP- number of false positive pixels for given classGT- number of ground truth pixels for given class
A blob with a BGR image in the format: [B, C=3, H=512, W=896], where:
- B – batch size
- C – number of channels
- H – image height
- W – image width
The output is a blob with shape [B, C=4, H=512, W=896]. It can be treated as a 4 channels feature map, where each feature map channel is a probability of one of the classes (BG, road, curb, mark).
[*] Other names and brands may be claimed as the property of others.