search

david8862 / keras-YOLOv3-model-set

end-to-end YOLOv4/v3/v2 object detection pipeline, implemented on tf.keras with different technologies
MIT License
640 stars 222 forks source link

yolov4 darknet benchmarking on PascalVOC2007 #107

Open nmochar2 opened 4 years ago

nmochar2 commented 4 years ago

Hi,

I'm having some difficult replicating results for Pascal VOC with the YoloV4 Darknet. I'm running with the command: python3 train.py --model_type=yolo4_darknet --anchors_path=configs/yolo4_anchors.txt --annotation_file=2007_train.txt --val_annotation_file=2007_val.txt --classes_path=configs/voc_classes --learning_rate=0.001 --label_smoothing=0.1 --enhance_augment=mosaic --optimizer=sgd --multi_anchor_assign --multiscale . I ran the same command for the YoloV4 Efficientnet model and can get comparable results to what you have. Were you able to get results similar to what was published in the YoloV4 paper or to your YoloV4 Efficientnet model and what command did you use?

Thank you

david8862 commented 4 years ago

Hi @nmochar2 , for the darknet backbone YOLO families, now this repo couldn't achieve a good training performance due to lack of valid Imagenet pretained backbone weights. I'm still trying to plan for resolving this problem. Sorry for trouble

nmochar2 commented 4 years ago

Oh okay. Thanks! I wasn't sure if I was doing something wrong and wanted to double check. Thanks for everything else