ncnn project
This repo is dedicated to the my ncnn's toy projects.
content
Gated-SCNN
The heaviest model best_cityscapes_checkpoint.pth
are used.
The input size is fixed at h, w = 512, 1024
.
The .param
and .bin
can be taken from google drive
Take a random sample from a Beijing street scene
CenterNet
The model resnet50_coco_wo_deconv.pth
are used.
The input size is fixed at h, w = 448, 672
.
The ncnn weight google drive
UNet-VB
The model is just a very common UNet with ResNet
The input size is fixed at h, w = 256, 256
. In order to achieve a better effect on high resolution pictures, the grid cutting trick was used.
The ncnn weight google drive
FaceMesh
Not exactly the same as the google official implementation.
The main difference is that the face detector in MediaPipe is not used here. Besides, the five keypoints of the face were not detected. So the face is not aligned horizontally and going straight to the Landmarks detection. Obviously, this must make a performance difference.