This is a segmentation network to classify each pixel into 20 classes:
Metric | Value |
---|---|
Image size | 2048x1024 |
GFlops | 58.572 |
MParams | 6.686 |
Source framework | Caffe* |
The quality metrics calculated on 2000 images:
Label | IOU |
---|---|
mean | 0.6907 |
Road | 0.910379 |
Sidewalk | 0.630676 |
Building | 0.860139 |
Wall | 0.424166 |
Fence | 0.592632 |
Pole | 0.559078 |
Traffic Light | 0.654779 |
Traffic Sign | 0.648217 |
Vegetation | 0.882593 |
Terrain | 0.620521 |
Sky | 0.976889 |
Person | 0.711653 |
Rider | 0.612787 |
Car | 0.877892 |
Truck | 0.674829 |
Bus | 0.743752 |
Train | 0.358641 |
Motorcycle | 0.600701 |
Bicycle | 0.622246 |
Ego-Vehicle | 0.852932 |
IOU=TP/(TP+FN+FP)
, where:TP
- number of true positive pixels for given classFN
- number of false negative pixels for given classFP
- number of false positive pixels for given classThe blob with BGR image in format: [B, C=3, H=1024, W=2048], where:
[*] Other names and brands may be claimed as the property of others.