search

TNTWEN / OpenVINO-YOLOV4

This is implementation of YOLOv4,YOLOv4-relu,YOLOv4-tiny,YOLOv4-tiny-3l,Scaled-YOLOv4 and INT8 Quantization in OpenVINO2021.3
MIT License
238 stars 66 forks source link

Yolov4-Tiny: No detections when object is small or large #51

Open at1993 opened 3 years ago

at1993 commented 3 years ago

Hi TNTWEN,

I recently converted a custom yolov4-tiny model and have it converted to IR using this script. I then did a side by side comparison between the darknet model and the converted IR mode. I noticed that the converted model is having trouble detecting my object when I bring it closer to the camera or further away from the the camera. Meanwhile the darknet model has no trouble detect it at all. Do you have any idea why this is the case?

at1993 commented 3 years ago

I think I found my problem, I did not change the number of class in the yolov4-tiny.json file. After I corrected that, it seems to perform better.

at1993 commented 3 years ago

Note: the masks values in the yolo_v4_tiny.json should match the one used in the cfg file. For custom yolov4-tiny, the following should be changed: "masks": [[3, 4, 5], [1, 2, 3]], to "masks": [[3, 4, 5], [0,1, 2]], Otherwise, the bounding box will be big when the object becomes small

el-stinjo commented 2 years ago

@at1993 Thanks! works like a charm :)