search

thtrieu / darkflow

Translate darknet to tensorflow. Load trained weights, retrain/fine-tune using tensorflow, export constant graph def to mobile devices
GNU General Public License v3.0
6.13k stars 2.08k forks source link

darkflow results differ from darknet #718

Open glenn-jocher opened 6 years ago

glenn-jocher commented 6 years ago

I see different results when running darkflow and darknet on the same image, using the same default .weight and .cfg files.

Darknet (black) appears more accurate than darkflow (white). The dog is notably missed in both images by darkflow, but successfully detected by darknet. I've made no changes to the default *.cfg file, threshold is 0.6, input dimension is 416.

Darknet confidences are generally higher with better bounding boxes as well.

/darknet/data/dog.jpg with yolov2-tiny.weights and /darknet/cfg/yolov2-tiny.cfg dog yolov2-tiny

/darknet/data/person.jpg with yolov2-tiny.weights and /darknet/cfg/yolov2-tiny.cfg person yolov2-tiny

lfb94 commented 6 years ago

Hello,

Did you figure this out?

Thanks

wns349 commented 6 years ago

Same problem here. I changed weights_walker's starting offset from 16 to 20 to get it running with yolov2-tiny.weights. I am assuming that's not the problem here though.

glenn-jocher commented 6 years ago

@wns349 yes the starting offset must be changed from 16 to 20 before running yolov2-tiny. I don't think this is the issue though, as I also see difference in yolov2 with the default offset of 16. I think there might be structural inconsistencies in the darkflow implementation that cause it to differ.

Malouke commented 6 years ago

i am real disappointed about darkflow i spend 3 weeks on labelling and training data but without result even with million images of cells and annotations really not good