we implement two versions of the model, one of which is a real-time high-precision version on GPU platform and the other is a real-time version on CPU platform.
model | input size | Val iou (Supervisely) | Val iou (Whole) | speed (FPS) | Size (MB) |
---|---|---|---|---|---|
Linknet_res18 | 640 | 88% | 78% | 100 (P40) | 44 |
Unet++ | 256 | 56% | --- | 24 (CPU) | 1.58 |
Dataset | Train | Val |
---|---|---|
Supervisely | 4757 | 526 |
VOC_person | 401 | 44 |
COCO_person | 58023 | 6411 |
COCO_negtive | 54172 | 0 |
VIP | 16702 | 1846 |
TRIMODAL | 5722 | 0 |
Total | 139777 | 8827 |
1). Linknet
we use linknet_res18 as the model to implement the real-time high-precision version on GPU platform
2). Unet++
we use unet++ as the model to implement the real-time version on CPU platform. Notice that the model is not the same as the origin unet++ mentioned in paper. while testing , we only keep X^00 , X^01 , X^10 to do the inference .
download the pretrained linknet model[BaiduNet](code: ftfe), and run the command below to try the demo
python ./demo/demo.py