search

hunglc007 / tensorflow-yolov4-tflite

YOLOv4, YOLOv4-tiny, YOLOv3, YOLOv3-tiny Implemented in Tensorflow 2.0, Android. Convert YOLO v4 .weights tensorflow, tensorrt and tflite
https://github.com/hunglc007/tensorflow-yolov4-tflite
MIT License
2.24k stars 1.24k forks source link

why tf-yolov4 has large different result with darknet? #38

Open liuxufenfeiya opened 4 years ago

liuxufenfeiya commented 4 years ago

this is darknet result darknet416

and this is yolov4-tf yolov4-tf

i use my own trained .weights to infer that image, both input_size are 416, thresh is 0.5,iou_thresh is 0.45, but i don't konw why tf-yolov4 has large different result with darknet?

here is darknet results: darknet_result

hunglc007 commented 4 years ago

u can change thresh and iou_thresh in detect.py line 89

liuxufenfeiya commented 4 years ago

u can change thresh and iou_thresh in detect.py line 89

thank you for reply 1.thresh is in line 89,but iou_thresh is in line 90. 2.i try the same thresh and iou_thresh in darknet and tf-yolov4,but the result has large difference,i wonder why(thresh=0.5,iou_thresh=0.45)

liuxufenfeiya commented 4 years ago

i also tested the pytorch yolov4,the result is the nearing darknet here is pytorch result(thresh=0.5,iou_thresh=0.45) predictions

titanbender commented 4 years ago

I've had similar experiences with converting darknet models to TF. Not sure why that is, but there seems to be a large performance drop.