After training custom data in YOLOV8 Image segmentation it gives output float32[1,37,8400] and float32[1,160,160,32] where one is prediction and another is detection image edges. But in android there have no way to perform non maximum suppression to get single value for the detected image with confidence. Is there any way to stich NMS with ONNX model when converting YOLOV8 model to ONNX model(If able to add NMS in ONNX then may be can convert it in TFLite). If we can't do that with YOLOV8 then what is the best way to train custom dataset for image segmentation which will support in mobile devices.
Asked
Active
Viewed 65 times
0
Sudip Sadhukhan
- 1,784
- 12
- 21
-
This discussion: https://github.com/ultralytics/ultralytics/issues/4427 assumes that you can try yolov5 with further Tensorflow and TFLite conversion using this repository: https://github.com/AarohiSingla/TFLite-Object-Detection-Android-App-Tutorial-Using-YOLOv5/tree/main. It is for object detection, but yolov5 supports segmentation task too. Hope it will help – hanna_liavoshka Aug 24 '23 at 19:37