YOLACT official code address: https://github.com/dbolya/yolact/
YOLACT：https://arxiv.org/abs/1904.02689 YOLACT++：https://arxiv.org/abs/1912.06218 DCN：https://arxiv.org/abs/1703.06211
Pytorch convert to ONNX:
github: https://github.com/Ma-Dan/yolact/tree/onnx
CSDN: https://blog.csdn.net/qq_37532213/article/details/121671389
ZhiHu: https://zhuanlan.zhihu.com/p/128974102
**Note: We only implemented the forward reasoning of the YOLACT onnx model with DarkNet53 as the backbone. Other models are similar and can be modified and applied appropriately.
**The forward reasoning code completely abandons the dependence on torch and tensor, and uses numpy to express the data completely.**In addition, since YOLACT++ uses the DCNv2 operator, and onnx does not support dynamic convolution operators, it cannot be converted to onnx.
**If you must convert it, you can try using OpenCV's dynamic convolution or use normal convolution instead.
How our project uses:
Please download yolact_darknet53.onnx from Baidu network disk and place it in the ./model directory.
Baidu network disk address: https://pan.baidu.com/s/1r8Fss-IuxJ9yBDQxYUf3Rg, extraction code: koex
#Install dependency environment, we use python 3.7.11
pip install -r requriements.txt#View help instructions
python yolact_onnx_detect.py -h
usage: yolact_onnx_detect.py [-h] [--images IMAGES] [--image IMAGE]
[--onnx ONNX] [--dst DST] [--net_w NET_W]
[--net_h NET_H] [--conf_thr CONF_THR]
[--nms_thr NMS_THR] [--top_k TOP_K]
optional arguments:
-h, --help show this help message and exit
--images IMAGES input images path
--image IMAGE input image file
--onnx ONNX onnx model file
--dst DST detection results save path
--net_w NET_W networks input width
--net_h NET_H networks input height
--conf_thr CONF_THR detect confidence thresh
--nms_thr NMS_THR detection fast nms thresh
--top_k TOP_K max detection object number#execute script
python yolact_onnx_detect.py --images ./img/ --onnx ./model/yolact_darknet53.onnx --dst ./rslts/Visualization of inspection results：
