Object detection yolo v4 support
hcdiekmann opened this issue · 2 comments
hcdiekmann commented
The default supplied object detection model (ssd_mobilenet_v2_coco) has a bad precision of >25% mAP.
However after testing the object detection node with multiple newer models converted to IR format, for example yolo-v4-tf, it seems that models with more than one output layer are not supported and cause an error.
LewisLiuPub commented
Recently there are several requirements from real user/customers for supporting Yolo v7. sounds now it is the good time to enable it.
Get the model info from the openvino nootbooks.
The possible steps to complete the enabling task:
- Get the YOLOV7 Pytorch pre-trained files.
- Reference the steps to convert yolov5 model to OpenVINO optimized model.
- Either: Try to update model class for Yolov5 to support both Yolov5 and yolov7.
- Or if the existing class for yolov5 model doesn't support yolov7:
- Add new sub-class as child of base model for Yolov7 model parsing.
- Update Output ImageWindow to support the new model.
tmcourie commented
ROS Team 1 working on this